{"text": "大家好,最近一個名為Dark Bloom的項目,熱度驟增。它可以把許多閒置的Mac設備連成一個算力網絡。有人就把一台M5Mac 128GB內存的Mac進入了Dark Bloom,24小時內處理了13000多個請求,賺到了4.11美元,並且體現成功。至於電費,大約賺到收入的5%以及磨損其實都構不成問題。也就是說,在你桌上的Mac不用的時候,它可以替別人跑AI,順帶賺點錢。至少前沿模型的訂閱費就有了。這個由Icon Labs推出的分布式AI推理網絡,機制大概有這幾步。首先,開發者發起請求,Dark Bloom的調度中心負責挑一台合適的Mac,把這個請求加密發過去。Mac本地用MLX運行開運模型,再將結果加密返回。這個項目疊加了Secure Enclave,Secure Boot,還有設備認證,簽名程序,內存隔離等Apple的安全機制。嚴格來說,雖然中央調度方仍然存在信任邊界,不是絕對的端了端加密,但它比一般重包算力認真的多。Dark Bloom的成長速度非常快,今年4月上線,6月介入OpenRouter,7月初單日Token突破10億,7月底重購MLX的連續Batching Prefail速度提升3-4倍,到8月成為了OpenRouter的付費Provider。今天最新數據顯示,Dark Bloom已經介入1380台機器,網絡聚合內存123.2TB,AR2大約是22.2萬美元,實現一周翻倍。當然它的提升空間還是很大的,比如峰值利用率為15%,說明願意賣算力的機器比較多,而真實付費需求還沒有完全跟上。另外,收益分佈其實並不均勻,比如前面說的那台M5Max,它一天是賺了4.11美元,同時還有另一台M324GB的Mac,是在8個小時內一個任務都沒有接到的。所以也許調度它更偏向高內存,認證狀態好的機器。再者,它的延遲、穩定性還有知識模型數量等,勢必會繼續完善。上周末,剛看到這個項目的時候,想起幾年前有研究過分佈式計算,Distributed Computing,而YouTuber Matthew Burman就這樣找到,Dart Bloom稱為數據中心殺手。他說,Just imagine we have a million of these nodes set up,That is major inference power that anybody can tap into at a fraction of the price of going to an open AI or an anthropic,and it's completely distributed,it's completely built by regular people。當推理需求暴漲,而開源模型、MOE和電話技術等,又正好把一部分能力壓到了Mac、工作站還有AIPC這類邊緣設備上,那麼原本散落在千家萬戶的沉睡算力,就有了被組織成公共基礎設施的可能。數據中心仍會守住訓練和高并發推理,但那些隱私敏感突發的常委的成本敏感的推理需求,完全可以流向一張有可驗證邊緣設備組成的網路。最後介紹一下DarkGloom的核心開發者Gajesh,他7歲開始寫代碼,10歲給創業公司做聊天機器人,13歲創建了一個基於區塊鏈的加密貨幣理財管理工具,資產管理規模一度達到700萬美元,15歲加入EigenLabs,今年Gajesh干滿18歲,致力於打造開放網絡,加速開放智能。這期節目就到這裡,Thanks for listening!我是Kyoli,Thanks for listening!", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 29.560000000000002, "text": "大家好,最近一個名為Dark Bloom的項目,熱度驟增。它可以把許多閒置的Mac設備連成一個算力網絡。有人就把一台M5Mac 128GB內存的Mac進入了Dark Bloom,24小時內處理了13000多個請求,賺到了4.11美元,並且體現成功。至於電費,大約賺到收入的5%以及磨損其實都構不成問題。也就是說,在你桌上的Mac不用的時候,它可以替別人跑AI,順帶賺點錢。", "tokens": [50365, 15248, 11, 46777, 8990, 15940, 6344, 35, 809, 25927, 1546, 8313, 227, 11386, 11, 32745, 13127, 24023, 253, 24228, 252, 1543, 11284, 6723, 16075, 40914, 6392, 5196, 240, 34719, 1546, 34355, 39035, 20854, 36252, 11336, 8990, 19497, 13486, 26352, 6948, 94, 1543, 45820, 3111, 16075, 2257, 15433, 44, 20, 34355, 29810, 8769, 28472, 39781, 1546, 34355, 18214, 14028, 2289, 35, 809, 25927, 11, 7911, 46997, 28472, 39289, 13876, 2289, 7668, 1360, 6392, 3338, 16302, 32718, 11, 13352, 118, 21381, 19, 13, 5348, 39433, 11, 35934, 20334, 23987, 9581, 11336, 32311, 1543, 20844, 19488, 20545, 50071, 11, 3582, 30889, 13352, 118, 4511, 18681, 14028, 1546, 20, 4, 40282, 163, 13033, 22220, 235, 14139, 7182, 43362, 1960, 11336, 17197, 1543, 6404, 26222, 11, 3581, 2166, 162, 5475, 5708, 1546, 34355, 24384, 20643, 11, 11284, 6723, 9531, 123, 16158, 4035, 32585, 48698, 11, 8313, 228, 32379, 13352, 118, 8216, 29483, 1543, 51843], "temperature": 0, "avg_logprob": -0.15426332072207802, "compression_ratio": 1.2078651685393258, "no_speech_prob": 8.688429316283663e-12}, {"id": 1, "seek": 2956, "start": 29.56, "end": 54.56, "text": "至少前沿模型的訂閱費就有了。這個由Icon Labs推出的分布式AI推理網絡,機制大概有這幾步。首先,開發者發起請求,Dark Bloom的調度中心負責挑一台合適的Mac,把這個請求加密發過去。Mac本地用MLX運行開運模型,再將結果加密返回。這個項目疊加了Secure Enclave,Secure Boot,還有設備認證,簽名程序,內存隔離等Apple的安全機制。", "tokens": [50365, 20844, 15686, 8945, 3308, 123, 41908, 39823, 1546, 30215, 50071, 3111, 2412, 2289, 1543, 6287, 23786, 40, 1671, 40047, 33597, 7781, 1546, 6627, 34688, 27584, 48698, 33597, 13876, 26352, 6948, 94, 11, 17543, 25491, 32044, 2412, 2664, 23575, 31429, 1543, 36490, 11, 8949, 14637, 12444, 14637, 9147, 16302, 32718, 11, 35, 809, 25927, 1546, 32943, 13127, 5975, 7945, 11561, 254, 11561, 105, 49067, 2257, 15433, 14245, 3330, 102, 1546, 34355, 11, 16075, 6287, 16302, 32718, 9990, 47838, 14637, 36910, 1543, 34355, 8802, 10928, 9254, 12683, 55, 32772, 8082, 8949, 32772, 41908, 39823, 11, 8623, 34168, 35181, 9990, 47838, 3316, 242, 8350, 1543, 6287, 8313, 227, 11386, 18140, 232, 9990, 2289, 29511, 540, 29584, 27995, 11, 29511, 540, 37263, 11, 15569, 39035, 20854, 22041, 11823, 231, 11, 24579, 121, 15940, 29649, 6346, 237, 11, 28472, 39781, 165, 1206, 32662, 10187, 9132, 306, 1546, 45466, 17543, 25491, 1543, 51615], "temperature": 0, "avg_logprob": -0.12914863586425782, "compression_ratio": 1.186246418338109, "no_speech_prob": 1.4848459267691183e-11}, {"id": 2, "seek": 5456, "start": 54.56, "end": 62.56, "text": "嚴格來說,雖然中央調度方仍然存在信任邊界,不是絕對的端了端加密,但它比一般重包算力認真的多。", "tokens": [50365, 29593, 112, 30921, 36979, 11, 6306, 244, 5823, 5975, 1787, 106, 32943, 13127, 9249, 1550, 235, 5823, 39781, 3581, 17665, 26443, 14901, 21185, 11, 7296, 6948, 243, 2855, 1546, 11957, 107, 2289, 11957, 107, 9990, 47838, 11, 8395, 11284, 11706, 2257, 49640, 12624, 23305, 19497, 13486, 22041, 8034, 6392, 1543, 50765], "temperature": 0, "avg_logprob": -0.10623514118479259, "compression_ratio": 1.085987261146497, "no_speech_prob": 1.669749755128791e-11}, {"id": 3, "seek": 5456, "start": 62.56, "end": 77.56, "text": "Dark Bloom的成長速度非常快,今年4月上線,6月介入OpenRouter,7月初單日Token突破10億,7月底重購MLX的連續Batching Prefail速度提升3-4倍,到8月成為了OpenRouter的付費Provider。", "tokens": [50765, 35, 809, 25927, 1546, 11336, 15353, 31217, 13127, 14392, 10251, 11, 38937, 19, 6939, 5708, 31699, 11, 21, 6939, 30312, 14028, 45569, 49, 23985, 11, 22, 6939, 28727, 24227, 6890, 51, 8406, 40859, 39560, 3279, 25459, 11, 22, 6939, 22816, 12624, 13352, 120, 12683, 55, 1546, 36252, 29583, 33, 29569, 6001, 69, 864, 31217, 13127, 20949, 41670, 18, 12, 19, 35477, 11, 4511, 23, 6939, 11336, 6344, 2289, 45569, 49, 23985, 1546, 41469, 50071, 12681, 85, 1438, 1543, 51515], "temperature": 0, "avg_logprob": -0.10623514118479259, "compression_ratio": 1.085987261146497, "no_speech_prob": 1.669749755128791e-11}, {"id": 4, "seek": 7756, "start": 77.56, "end": 88.56, "text": "今天最新數據顯示,Dark Bloom已經介入1380台機器,網絡聚合內存123.2TB,AR2大約是22.2萬美元,實現一周翻倍。", "tokens": [50365, 12074, 8661, 12560, 30622, 36841, 46605, 25696, 11, 35, 809, 25927, 25338, 30312, 14028, 7668, 4702, 15433, 17543, 34386, 11, 26352, 6948, 94, 8171, 248, 14245, 28472, 39781, 4762, 18, 13, 17, 51, 33, 11, 1899, 17, 3582, 30889, 1541, 7490, 13, 17, 14099, 39433, 11, 10376, 9581, 2257, 37390, 42716, 35477, 1543, 50915], "temperature": 0, "avg_logprob": -0.08500791045854676, "compression_ratio": 1.0629629629629629, "no_speech_prob": 1.8627736506071635e-11}, {"id": 5, "seek": 7756, "start": 88.56, "end": 97.56, "text": "當然它的提升空間還是很大的,比如峰值利用率為15%,說明願意賣算力的機器比較多,而真實付費需求還沒有完全跟上。", "tokens": [50915, 21707, 45224, 20949, 41670, 23322, 11016, 25583, 4563, 39156, 11, 36757, 26555, 108, 40242, 23700, 9254, 44866, 6344, 5211, 8923, 4622, 11100, 20692, 9042, 13352, 96, 19497, 13486, 1546, 17543, 34386, 25174, 6392, 11, 11070, 6303, 10376, 41469, 50071, 32535, 32718, 7824, 6963, 37100, 9678, 5708, 1543, 51365], "temperature": 0, "avg_logprob": -0.08500791045854676, "compression_ratio": 1.0629629629629629, "no_speech_prob": 1.8627736506071635e-11}, {"id": 6, "seek": 9756, "start": 97.56, "end": 108.56, "text": "另外,收益分佈其實並不均勻,比如前面說的那台M5Max,它一天是賺了4.11美元,同時還有另一台M324GB的Mac,是在8個小時內一個任務都沒有接到的。", "tokens": [50365, 26202, 11, 18681, 5419, 232, 6627, 1593, 230, 14139, 35934, 1960, 14872, 229, 7978, 119, 11, 36757, 8945, 8833, 4622, 1546, 4184, 15433, 44, 20, 36025, 11, 11284, 2257, 6135, 1541, 13352, 118, 2289, 19, 13, 5348, 39433, 11, 13089, 6611, 15569, 22762, 2257, 15433, 44, 18, 7911, 8769, 1546, 34355, 11, 1541, 3581, 23, 3338, 46997, 28472, 8990, 26443, 39517, 7182, 6963, 14468, 4511, 1546, 1543, 50915], "temperature": 0, "avg_logprob": -0.16759314007229276, "compression_ratio": 1.1371158392434988, "no_speech_prob": 1.8087035341673996e-11}, {"id": 7, "seek": 9756, "start": 108.56, "end": 112.56, "text": "所以也許調度它更偏向高內存,認證狀態好的機器。", "tokens": [50915, 7239, 6404, 40914, 32943, 13127, 11284, 19002, 7437, 237, 24282, 12979, 28472, 39781, 11, 22041, 11823, 231, 18637, 222, 45248, 20715, 17543, 34386, 1543, 51115], "temperature": 0, "avg_logprob": -0.16759314007229276, "compression_ratio": 1.1371158392434988, "no_speech_prob": 1.8087035341673996e-11}, {"id": 8, "seek": 9756, "start": 112.56, "end": 118.56, "text": "再者,它的延遲、穩定性還有知識模型數量等,勢必會繼續完善。", "tokens": [51115, 8623, 12444, 11, 45224, 17527, 114, 3330, 110, 1231, 13740, 102, 12088, 21686, 15569, 6498, 43143, 41908, 39823, 30622, 26748, 10187, 11, 7978, 95, 28531, 6236, 38459, 14128, 5234, 226, 1543, 51415], "temperature": 0, "avg_logprob": -0.16759314007229276, "compression_ratio": 1.1371158392434988, "no_speech_prob": 1.8087035341673996e-11}, {"id": 9, "seek": 9756, "start": 118.56, "end": 126.56, "text": "上周末,剛看到這個項目的時候,想起幾年前有研究過分佈式計算,Distributed Computing,而YouTuber Matthew Burman就這樣找到,", "tokens": [51415, 5708, 37390, 1474, 104, 11, 16940, 18032, 6287, 8313, 227, 11386, 20643, 11, 7093, 9147, 23575, 49947, 2412, 23230, 242, 44704, 8816, 6627, 1593, 230, 27584, 30114, 19497, 11, 35, 468, 2024, 4866, 37804, 278, 11, 11070, 3223, 51, 10261, 12434, 7031, 1601, 3111, 8377, 25085, 4511, 11, 51815], "temperature": 0, "avg_logprob": -0.16759314007229276, "compression_ratio": 1.1371158392434988, "no_speech_prob": 1.8087035341673996e-11}, {"id": 10, "seek": 12656, "start": 126.56, "end": 129.56, "text": "Dart Bloom稱為數據中心殺手。", "tokens": [50365, 35, 446, 25927, 49606, 6344, 30622, 36841, 5975, 7945, 45528, 11389, 1543, 50515], "temperature": 0, "avg_logprob": -0.19474613132761487, "compression_ratio": 1.1805929919137466, "no_speech_prob": 4.447453516576161e-11}, {"id": 11, "seek": 12656, "start": 129.56, "end": 130.56, "text": "他說,", "tokens": [50515, 31592, 11, 50565], "temperature": 0, "avg_logprob": -0.19474613132761487, "compression_ratio": 1.1805929919137466, "no_speech_prob": 4.447453516576161e-11}, {"id": 12, "seek": 12656, "start": 130.56, "end": 133.56, "text": "Just imagine we have a million of these nodes set up,", "tokens": [50565, 12794, 3811, 321, 362, 257, 2459, 295, 613, 13891, 992, 493, 11, 50715], "temperature": 0, "avg_logprob": -0.19474613132761487, "compression_ratio": 1.1805929919137466, "no_speech_prob": 4.447453516576161e-11}, {"id": 13, "seek": 12656, "start": 133.56, "end": 140.56, "text": "That is major inference power that anybody can tap into at a fraction of the price of going to an open AI or an anthropic,", "tokens": [50715, 6390, 307, 2563, 38253, 1347, 300, 4472, 393, 5119, 666, 412, 257, 14135, 295, 264, 3218, 295, 516, 281, 364, 1269, 7318, 420, 364, 22727, 299, 11, 51065], "temperature": 0, "avg_logprob": -0.19474613132761487, "compression_ratio": 1.1805929919137466, "no_speech_prob": 4.447453516576161e-11}, {"id": 14, "seek": 12656, "start": 140.56, "end": 144.56, "text": "and it's completely distributed,it's completely built by regular people。", "tokens": [51065, 474, 309, 311, 2584, 12631, 11, 270, 311, 2584, 3094, 538, 3890, 561, 1543, 51265], "temperature": 0, "avg_logprob": -0.19474613132761487, "compression_ratio": 1.1805929919137466, "no_speech_prob": 4.447453516576161e-11}, {"id": 15, "seek": 12656, "start": 144.56, "end": 153.56, "text": "當推理需求暴漲,而開源模型、MOE和電話技術等,又正好把一部分能力壓到了Mac、工作站還有AIPC這類邊緣設備上,", "tokens": [51265, 13118, 33597, 13876, 32535, 32718, 23826, 112, 14065, 110, 11, 11070, 8949, 47402, 41908, 39823, 1231, 18976, 36, 12565, 20545, 11103, 32502, 9890, 241, 10187, 11, 17047, 15789, 2131, 16075, 2257, 32174, 8225, 13486, 45964, 21381, 34355, 1231, 41315, 34155, 15569, 32, 9139, 34, 2664, 40092, 14901, 12608, 96, 39035, 20854, 5708, 11, 51715], "temperature": 0, "avg_logprob": -0.19474613132761487, "compression_ratio": 1.1805929919137466, "no_speech_prob": 4.447453516576161e-11}, {"id": 16, "seek": 15356, "start": 153.56, "end": 159.56, "text": "那麼原本散落在千家萬戶的沉睡算力,就有了被組織成公共基礎設施的可能。", "tokens": [50365, 20504, 19683, 8802, 7017, 96, 30848, 3581, 20787, 5155, 14099, 1486, 114, 1546, 3308, 231, 40490, 19497, 13486, 11, 3111, 2412, 2289, 23238, 30052, 23141, 242, 11336, 13545, 15408, 26008, 14971, 236, 39035, 4307, 121, 1546, 16657, 1543, 50665], "temperature": 0, "avg_logprob": -0.10341348057299589, "compression_ratio": 1.184873949579832, "no_speech_prob": 3.154735153665378e-11}, {"id": 17, "seek": 15356, "start": 159.56, "end": 170.56, "text": "數據中心仍會守住訓練和高并發推理,但那些隱私敏感突發的常委的成本敏感的推理需求,完全可以流向一張有可驗證邊緣設備組成的網路。", "tokens": [50665, 30622, 36841, 5975, 7945, 1550, 235, 6236, 48480, 21632, 5396, 241, 47027, 12565, 12979, 3509, 114, 14637, 33597, 13876, 11, 8395, 4184, 13824, 10673, 109, 20083, 7017, 237, 9709, 40859, 14637, 1546, 11279, 44230, 1546, 11336, 8802, 7017, 237, 9709, 1546, 33597, 13876, 32535, 32718, 11, 37100, 6723, 27854, 24282, 2257, 18889, 2412, 4429, 24023, 245, 11823, 231, 14901, 12608, 96, 39035, 20854, 30052, 11336, 1546, 26352, 24658, 1543, 51215], "temperature": 0, "avg_logprob": -0.10341348057299589, "compression_ratio": 1.184873949579832, "no_speech_prob": 3.154735153665378e-11}, {"id": 18, "seek": 17056, "start": 170.56, "end": 177.56, "text": "最後介紹一下DarkGloom的核心開發者Gajesh,他7歲開始寫代碼,10歲給創業公司做聊天機器人,", "tokens": [50365, 20578, 30312, 38193, 8861, 35, 809, 38, 752, 298, 1546, 8987, 116, 7945, 8949, 14637, 12444, 38, 1805, 14935, 11, 5000, 22, 26256, 21017, 4510, 104, 19105, 16337, 120, 11, 3279, 26256, 17798, 5935, 113, 27119, 13545, 32981, 10907, 40096, 6135, 17543, 34386, 4035, 11, 50715], "temperature": 0, "avg_logprob": -0.15280311847555225, "compression_ratio": 1.1226993865030674, "no_speech_prob": 2.7072077218859292e-11}, {"id": 19, "seek": 17056, "start": 177.56, "end": 186.56, "text": "13歲創建了一個基於區塊鏈的加密貨幣理財管理工具,資產管理規模一度達到700萬美元,15歲加入EigenLabs,", "tokens": [50715, 7668, 26256, 5935, 113, 34157, 2289, 8990, 26008, 19488, 44526, 26268, 46241, 230, 1546, 9990, 47838, 11561, 101, 3509, 96, 13876, 11561, 94, 23131, 13876, 23323, 39806, 11, 35380, 33299, 23131, 13876, 2888, 237, 41908, 2257, 13127, 33086, 4511, 18197, 14099, 39433, 11, 5211, 26256, 43878, 36, 3213, 43, 17243, 11, 51165], "temperature": 0, "avg_logprob": -0.15280311847555225, "compression_ratio": 1.1226993865030674, "no_speech_prob": 2.7072077218859292e-11}, {"id": 20, "seek": 17056, "start": 186.56, "end": 191.56, "text": "今年Gajesh干滿18歲,致力於打造開放網絡,加速開放智能。", "tokens": [51165, 38937, 38, 1805, 14935, 26111, 32197, 6494, 26256, 11, 6784, 112, 13486, 19488, 12467, 37583, 8949, 12744, 26352, 6948, 94, 11, 9990, 31217, 8949, 12744, 5094, 118, 8225, 1543, 51415], "temperature": 0, "avg_logprob": -0.15280311847555225, "compression_ratio": 1.1226993865030674, "no_speech_prob": 2.7072077218859292e-11}, {"id": 21, "seek": 17056, "start": 191.56, "end": 193.56, "text": "這期節目就到這裡,Thanks for listening!", "tokens": [51415, 2664, 16786, 42567, 45918, 18771, 11, 27917, 337, 4764, 0, 51515], "temperature": 0, "avg_logprob": -0.15280311847555225, "compression_ratio": 1.1226993865030674, "no_speech_prob": 2.7072077218859292e-11}, {"id": 22, "seek": 19356, "start": 193.56, "end": 195.56, "text": "我是Kyoli,Thanks for listening!", "tokens": [50365, 15914, 42, 88, 9384, 11, 27917, 337, 4764, 0, 50465], "temperature": 0, "avg_logprob": -1.0108482042948406, "compression_ratio": 0.8048780487804879, "no_speech_prob": 9.534304457048393e-10}], "language": "zh"}