{"text": "早上好呀,我是阿黎。歡迎收看8月24日第130期AI早報。內容根據公開信息整理,歡迎留言補充和糾錯。以下是今日Github熱榜。先看算力與基建領域的最新動態。阿黎擬配股籌資約102億美元,全額投入AI戰略,覆蓋芯片,基礎設施,模型及部署。尽管上季净利潤驟降75%,仍加碼算力,凸顯科技巨頭壓住AI長期優勢,但基礎設施競賽成本高昂,能否兌現回報成焦點。英伟達斥資60億美元授權PullSide AI技術並追加10億投資,估值達120億,超百名員工轉入共研Namotron開源模型,目標挑戰中美前沿、Aulama等夥伴加入合作。Meta每年斥資數億美元租用微軟Adger算力,每周消耗Token達數萬億,用於訓練和運行AI模型。此舉凸顯科技巨頭對雲端算力的依賴,也反映AI競賽中基礎設施投入的激烈競爭。英伟達AI服務器漲價超15%,GB300售價約60萬美元,明年初出貨,存儲芯片成本上升式主因,三星、SK、海力士等供應商推動,反映AI算力需求火熱正推高供應鏈價格。監管動態方面,美國FDA釋放重要信號。美國FDA擬按臨床醫生標準監管生成式AI醫療設備,而非傳統軟件。新規要求批准前評估任務可靠性,部署後持續監控,以公開徵求意見。行業動態持續更新,先看平台交易,再看前沿技術。AI開發平台Hugging Face正探索出售,潛在估值至少130億美元。相比2023年45億美元的估值大幅提升,公司已與銀行合作測試買家興趣,尚未達成交易。新加坡啟動活體人腦細胞服務器項目,采用幹細胞培養的百萬級神經元。該服務器功耗僅1000瓦,比傳統AI服務器低99%,瞄準藥物開發機器人及網絡安全領域。接下來關注安全與隱私方面的最新動態。MacOS版Codex被曝繞過TCC權限,將ChatGPT歷史記錄存於本地目錄,可能間接讀取敏感內容。此前OpenAI Apple插件已能訪問全部iMessage並代發短信。專家警告此舉破壞端道端加密,用戶及聯繫人未同意數據被AI獲取,隱私邊界遭嚴重衝擊。OpenAI修復了ChatGPT Codex的速率限制問題,歸因於常繪畫圖像處理,低效及標題生成耗算力。受影響付費用戶額度將於今日全額重置,緩解IOS端Codex Remote限額問題。AI助手Instinct被曝服務條款允許永久且不可撤銷地使用。用戶郵件消息等數據訓練模型,並可轉授權第三方,引發隱私擔憂。批評者提醒用戶誤在未細讀條款時泄露敏感信息。開源生態持續升溫,模型與開源領域迎來多項新進展。Vercel數據顯示,開源模型Token份額兩月內從28.4%飙升至62%,首次超越B元模型的38%。這一增長發生在B元需求同步加速的背景下,並非萎縮所致,反映企業採用開源趨勢正加速。Cloudflare開源內部AI操作系統 Cloudflare OS基於Workers構建,整合上下文與系統。通過Gatekeepers限制AI寫入Isolate隔離敏感數據,構建安全殺箱,解決Agent數據訪問風險,為企業AI應用提供更安全環境。開源AI Agent Hermes以YOLO模式全自動入侵泰國財政部,掃描漏洞並竊取2012年人士記錄,攻擊日誌含585個文件及憑證。事件凸顯無人監管Agent的自動化攻擊風險,引發責任歸屬爭議,涉及操作者、開發者與模型。DeepMind推出Recirculation方法,無需重訓即可讓Transformer在推理時地規進化,通過預填充階段反饋激活實現動態狀態跟蹤。生成成成本保持平坦,串行工作集中在預填充階段,突破層數限制。以上就是今天的全部內容了,感謝收看,我們明天見。MING PAO CANADA // MING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTOMING PAO TORONTO", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 14.0, "text": "早上好呀,我是阿黎。歡迎收看8月24日第130期AI早報。內容根據公開信息整理,歡迎留言補充和糾錯。以下是今日Github熱榜。", "tokens": [50365, 21176, 5708, 2131, 15348, 11, 15914, 13594, 6173, 236, 1543, 21315, 34333, 23, 6939, 7911, 6890, 7536, 42438, 16786, 48698, 21176, 17803, 1543, 28472, 25750, 31337, 36841, 13545, 8949, 17665, 26460, 27662, 13876, 11, 21315, 43681, 8083, 250, 2347, 227, 12565, 15239, 122, 13133, 1543, 3588, 4438, 1541, 26259, 38, 355, 836, 32745, 22989, 250, 1543, 51065], "temperature": 0, "avg_logprob": -0.1261291652917862, "compression_ratio": 1.1245421245421245, "no_speech_prob": 9.742299848791713e-12}, {"id": 1, "seek": 0, "start": 17.0, "end": 28.0, "text": "先看算力與基建領域的最新動態。阿黎擬配股籌資約102億美元,全額投入AI戰略,覆蓋芯片,基礎設施,模型及部署。", "tokens": [51215, 10108, 4200, 19497, 13486, 34127, 26008, 34157, 8313, 246, 16262, 253, 1546, 8661, 12560, 12572, 45248, 1543, 13594, 6173, 236, 13167, 105, 38846, 14356, 94, 22113, 234, 35380, 30889, 3279, 17, 25459, 39433, 11, 11319, 10178, 235, 37094, 14028, 48698, 23209, 6904, 98, 11, 2888, 228, 39198, 233, 13778, 107, 16668, 11, 26008, 14971, 236, 39035, 4307, 121, 11, 41908, 39823, 25703, 13470, 16469, 110, 1543, 51765], "temperature": 0, "avg_logprob": -0.1261291652917862, "compression_ratio": 1.1245421245421245, "no_speech_prob": 9.742299848791713e-12}, {"id": 2, "seek": 2800, "start": 28.0, "end": 39.0, "text": "尽管上季净利潤驟降75%,仍加碼算力,凸顯科技巨頭壓住AI長期優勢,但基礎設施競賽成本高昂,能否兌現回報成焦點。", "tokens": [50365, 1530, 121, 23131, 5708, 8052, 96, 6336, 222, 23700, 35622, 97, 24023, 253, 47421, 11901, 8923, 1550, 235, 9990, 16337, 120, 19497, 13486, 11, 6336, 116, 46605, 42091, 32502, 5238, 101, 19921, 45964, 21632, 48698, 15353, 16786, 12827, 103, 7978, 95, 11, 8395, 26008, 14971, 236, 39035, 4307, 121, 11957, 114, 45668, 11336, 8802, 12979, 1431, 224, 11, 8225, 45334, 2347, 234, 9581, 8350, 17803, 11336, 4971, 99, 8216, 1543, 50915], "temperature": 0, "avg_logprob": -0.13589370328616474, "compression_ratio": 1.0791139240506329, "no_speech_prob": 1.2132532825614994e-11}, {"id": 3, "seek": 2800, "start": 40.0, "end": 53.0, "text": "英伟達斥資60億美元授權PullSide AI技術並追加10億投資,估值達120億,超百名員工轉入共研Namotron開源模型,目標挑戰中美前沿、Aulama等夥伴加入合作。", "tokens": [50965, 27869, 7384, 253, 33086, 4307, 98, 35380, 4550, 25459, 39433, 6900, 230, 40212, 47, 858, 50, 482, 7318, 32502, 9890, 241, 35934, 39947, 9990, 3279, 25459, 37094, 35380, 11, 7384, 108, 40242, 33086, 28829, 25459, 11, 19869, 31906, 15940, 20992, 23323, 32739, 14028, 15408, 23230, 242, 45, 335, 310, 2044, 8949, 47402, 41908, 39823, 11, 11386, 46299, 49067, 23209, 5975, 9175, 8945, 3308, 123, 1231, 32, 425, 2404, 10187, 1787, 98, 7384, 112, 43878, 14245, 11914, 1543, 51615], "temperature": 0, "avg_logprob": -0.13589370328616474, "compression_ratio": 1.0791139240506329, "no_speech_prob": 1.2132532825614994e-11}, {"id": 4, "seek": 5300, "start": 53.0, "end": 68.0, "text": "Meta每年斥資數億美元租用微軟Adger算力,每周消耗Token達數萬億,用於訓練和運行AI模型。此舉凸顯科技巨頭對雲端算力的依賴,也反映AI競賽中基礎設施投入的激烈競爭。", "tokens": [50365, 44, 7664, 23664, 5157, 4307, 98, 35380, 30622, 25459, 39433, 8204, 253, 9254, 39152, 14454, 253, 15830, 1321, 19497, 13486, 11, 23664, 37390, 28837, 4450, 245, 51, 8406, 33086, 30622, 14099, 25459, 11, 9254, 19488, 5396, 241, 47027, 12565, 32772, 8082, 48698, 41908, 39823, 1543, 17947, 42828, 6336, 116, 46605, 42091, 32502, 5238, 101, 19921, 2855, 6306, 110, 11957, 107, 19497, 13486, 1546, 3254, 251, 164, 7062, 11, 6404, 22138, 1431, 254, 48698, 11957, 114, 45668, 5975, 26008, 14971, 236, 39035, 4307, 121, 37094, 14028, 1546, 26373, 222, 23661, 230, 11957, 114, 8164, 255, 1543, 51115], "temperature": 0, "avg_logprob": -0.06262793685450699, "compression_ratio": 1.028169014084507, "no_speech_prob": 1.972104077263559e-11}, {"id": 5, "seek": 6800, "start": 68.0, "end": 84.0, "text": "英伟達AI服務器漲價超15%,GB300售價約60萬美元,明年初出貨,存儲芯片成本上升式主因,三星、SK、海力士等供應商推動,反映AI算力需求火熱正推高供應鏈價格。", "tokens": [50365, 27869, 7384, 253, 33086, 48698, 27408, 39517, 34386, 14065, 110, 45441, 19869, 5211, 8923, 8769, 12566, 17198, 106, 45441, 30889, 4550, 14099, 39433, 11, 11100, 5157, 28727, 7781, 11561, 101, 11, 39781, 12827, 110, 13778, 107, 16668, 11336, 8802, 5708, 41670, 27584, 13557, 8698, 11, 10960, 20682, 1231, 50, 42, 1231, 20078, 13486, 30337, 10187, 3254, 249, 20481, 45581, 33597, 12572, 11, 22138, 1431, 254, 48698, 19497, 13486, 32535, 32718, 26171, 32745, 15789, 33597, 12979, 3254, 249, 20481, 46241, 230, 45441, 30921, 1543, 51165], "temperature": 0, "avg_logprob": -0.10320233750617368, "compression_ratio": 1.0251256281407035, "no_speech_prob": 2.038185939468029e-11}, {"id": 6, "seek": 8400, "start": 84.0, "end": 88.0, "text": "監管動態方面,美國FDA釋放重要信號。", "tokens": [50365, 5419, 96, 23131, 12572, 45248, 39604, 11, 21228, 37, 7509, 5873, 233, 12744, 24928, 17665, 18616, 1543, 50565], "temperature": 0, "avg_logprob": -0.08295291232079574, "compression_ratio": 1.1269035532994924, "no_speech_prob": 2.1755685794544455e-11}, {"id": 7, "seek": 8400, "start": 88.0, "end": 100.0, "text": "美國FDA擬按臨床醫生標準監管生成式AI醫療設備,而非傳統軟件。新規要求批准前評估任務可靠性,部署後持續監控,以公開徵求意見。", "tokens": [50565, 21228, 37, 7509, 13167, 105, 26613, 6784, 101, 6346, 232, 48351, 8244, 46299, 28220, 5419, 96, 23131, 8244, 11336, 27584, 48698, 48351, 6651, 224, 39035, 20854, 11, 11070, 12107, 40852, 33725, 14454, 253, 20485, 1543, 12560, 2888, 237, 4275, 32718, 3416, 117, 6336, 228, 8945, 39115, 7384, 108, 26443, 39517, 4429, 5363, 254, 21686, 11, 13470, 16469, 110, 5661, 17694, 29583, 5419, 96, 48707, 11, 3588, 13545, 8949, 2172, 113, 32718, 9042, 10470, 1543, 51165], "temperature": 0, "avg_logprob": -0.08295291232079574, "compression_ratio": 1.1269035532994924, "no_speech_prob": 2.1755685794544455e-11}, {"id": 8, "seek": 10000, "start": 100.0, "end": 107.0, "text": "行業動態持續更新,先看平台交易,再看前沿技術。", "tokens": [50365, 8082, 27119, 12572, 45248, 17694, 29583, 19002, 12560, 11, 10108, 4200, 16716, 15433, 28455, 31962, 11, 8623, 4200, 8945, 3308, 123, 32502, 9890, 241, 1543, 50715], "temperature": 0, "avg_logprob": -0.0781034871834476, "compression_ratio": 1.162162162162162, "no_speech_prob": 2.2665929896859005e-11}, {"id": 9, "seek": 10000, "start": 107.0, "end": 120.0, "text": "AI開發平台Hugging Face正探索出售,潛在估值至少130億美元。相比2023年45億美元的估值大幅提升,公司已與銀行合作測試買家興趣,尚未達成交易。", "tokens": [50715, 48698, 8949, 14637, 16716, 15433, 39, 697, 3249, 4047, 15789, 6900, 95, 7732, 95, 7781, 17198, 106, 11, 35622, 249, 3581, 7384, 108, 40242, 20844, 15686, 42438, 25459, 39433, 1543, 15106, 11706, 2009, 9356, 5157, 8465, 25459, 39433, 1546, 7384, 108, 40242, 3582, 3509, 227, 20949, 41670, 11, 13545, 32981, 15701, 34127, 42041, 222, 8082, 14245, 11914, 9592, 105, 22099, 31001, 5155, 44089, 39835, 11, 1530, 248, 29954, 33086, 11336, 28455, 31962, 1543, 51365], "temperature": 0, "avg_logprob": -0.0781034871834476, "compression_ratio": 1.162162162162162, "no_speech_prob": 2.2665929896859005e-11}, {"id": 10, "seek": 10000, "start": 120.0, "end": 126.0, "text": "新加坡啟動活體人腦細胞服務器項目,采用幹細胞培養的百萬級神經元。", "tokens": [51365, 12560, 9990, 14872, 94, 3284, 253, 12572, 25956, 23987, 4035, 21184, 99, 41588, 19095, 252, 27408, 39517, 34386, 8313, 227, 11386, 11, 5873, 229, 9254, 29932, 41588, 19095, 252, 16262, 117, 22217, 232, 1546, 31906, 14099, 35315, 20432, 15145, 14812, 1543, 51665], "temperature": 0, "avg_logprob": -0.0781034871834476, "compression_ratio": 1.162162162162162, "no_speech_prob": 2.2665929896859005e-11}, {"id": 11, "seek": 12600, "start": 126.0, "end": 134.0, "text": "該服務器功耗僅1000瓦,比傳統AI服務器低99%,瞄準藥物開發機器人及網絡安全領域。", "tokens": [50365, 22033, 27408, 39517, 34386, 32311, 4450, 245, 9502, 227, 21199, 163, 241, 99, 11, 11706, 40852, 33725, 48698, 27408, 39517, 34386, 41377, 8494, 8923, 36733, 226, 28220, 21782, 98, 23516, 8949, 14637, 17543, 34386, 4035, 25703, 26352, 6948, 94, 45466, 8313, 246, 16262, 253, 1543, 50765], "temperature": 0, "avg_logprob": -0.10300378596529047, "compression_ratio": 1.0662460567823344, "no_speech_prob": 2.0776004180933505e-11}, {"id": 12, "seek": 12600, "start": 134.0, "end": 138.0, "text": "接下來關注安全與隱私方面的最新動態。", "tokens": [50765, 40012, 14899, 26432, 45466, 34127, 10673, 109, 20083, 39604, 1546, 8661, 12560, 12572, 45248, 1543, 50965], "temperature": 0, "avg_logprob": -0.10300378596529047, "compression_ratio": 1.0662460567823344, "no_speech_prob": 2.0776004180933505e-11}, {"id": 13, "seek": 12600, "start": 138.0, "end": 151.0, "text": "MacOS版Codex被曝繞過TCC權限,將ChatGPT歷史記錄存於本地目錄,可能間接讀取敏感內容。此前OpenAI Apple插件已能訪問全部iMessage並代發短信。", "tokens": [50965, 34355, 4367, 42096, 34, 1429, 87, 23238, 9531, 251, 23141, 252, 8816, 51, 11717, 40212, 43446, 11, 34168, 41683, 38, 47, 51, 4287, 115, 45399, 16958, 9567, 226, 39781, 19488, 8802, 10928, 11386, 9567, 226, 11, 16657, 11016, 14468, 7422, 222, 29436, 7017, 237, 9709, 28472, 25750, 1543, 17947, 8945, 45569, 48698, 6373, 11673, 240, 20485, 15701, 8225, 5396, 103, 11361, 38714, 72, 44, 442, 609, 35934, 19105, 14637, 5881, 255, 17665, 1543, 51615], "temperature": 0, "avg_logprob": -0.10300378596529047, "compression_ratio": 1.0662460567823344, "no_speech_prob": 2.0776004180933505e-11}, {"id": 14, "seek": 15100, "start": 151.0, "end": 160.0, "text": "專家警告此舉破壞端道端加密,用戶及聯繫人未同意數據被AI獲取,隱私邊界遭嚴重衝擊。", "tokens": [50365, 46072, 5155, 39499, 16846, 17947, 42828, 39560, 12829, 252, 11957, 107, 6025, 11957, 107, 9990, 47838, 11, 9254, 1486, 114, 25703, 43942, 23141, 104, 4035, 29954, 13089, 9042, 30622, 36841, 23238, 48698, 29057, 110, 29436, 11, 10673, 109, 20083, 14901, 21185, 3330, 255, 29593, 112, 12624, 9890, 251, 37813, 1543, 50815], "temperature": 0, "avg_logprob": -0.08832076593493739, "compression_ratio": 1.0774410774410774, "no_speech_prob": 2.7862011309220946e-11}, {"id": 15, "seek": 15100, "start": 160.0, "end": 168.0, "text": "OpenAI修復了ChatGPT Codex的速率限制問題,歸因於常繪畫圖像處理,低效及標題生成耗算力。", "tokens": [50815, 45569, 48698, 7792, 106, 2172, 102, 2289, 41683, 38, 47, 51, 15549, 87, 1546, 31217, 44866, 43446, 25491, 17197, 11, 4287, 116, 8698, 19488, 11279, 23141, 103, 43822, 2523, 244, 12760, 39289, 13876, 11, 41377, 43076, 25703, 46299, 14475, 8244, 11336, 4450, 245, 19497, 13486, 1543, 51215], "temperature": 0, "avg_logprob": -0.08832076593493739, "compression_ratio": 1.0774410774410774, "no_speech_prob": 2.7862011309220946e-11}, {"id": 16, "seek": 15100, "start": 168.0, "end": 175.0, "text": "受影響付費用戶額度將於今日全額重置,緩解IOS端Codex Remote限額問題。", "tokens": [51215, 23151, 16820, 13665, 123, 41469, 50071, 9254, 1486, 114, 10178, 235, 13127, 34168, 19488, 26259, 11319, 10178, 235, 12624, 34719, 11, 12608, 102, 17278, 40, 4367, 11957, 107, 34, 1429, 87, 44858, 43446, 10178, 235, 17197, 1543, 51565], "temperature": 0, "avg_logprob": -0.08832076593493739, "compression_ratio": 1.0774410774410774, "no_speech_prob": 2.7862011309220946e-11}, {"id": 17, "seek": 17500, "start": 175.0, "end": 181.0, "text": "AI助手Instinct被曝服務條款允許永久且不可撤銷地使用。", "tokens": [50365, 48698, 37618, 11389, 28411, 5460, 23238, 9531, 251, 27408, 39517, 31106, 48798, 2347, 223, 40914, 42503, 25320, 20334, 1960, 4429, 20559, 97, 42041, 115, 10928, 22982, 9254, 1543, 50665], "temperature": 0, "avg_logprob": -0.08200719702334805, "compression_ratio": 1.138996138996139, "no_speech_prob": 2.7547306449826614e-11}, {"id": 18, "seek": 17500, "start": 181.0, "end": 187.0, "text": "用戶郵件消息等數據訓練模型,並可轉授權第三方,引發隱私擔憂。", "tokens": [50665, 9254, 1486, 114, 11178, 113, 20485, 28837, 26460, 10187, 30622, 36841, 5396, 241, 47027, 41908, 39823, 11, 35934, 4429, 32739, 6900, 230, 40212, 35878, 9249, 11, 41301, 14637, 10673, 109, 20083, 13167, 242, 29713, 224, 1543, 50965], "temperature": 0, "avg_logprob": -0.08200719702334805, "compression_ratio": 1.138996138996139, "no_speech_prob": 2.7547306449826614e-11}, {"id": 19, "seek": 17500, "start": 187.0, "end": 191.0, "text": "批評者提醒用戶誤在未細讀條款時泄露敏感信息。", "tokens": [50965, 3416, 117, 39115, 12444, 20949, 40796, 9254, 1486, 114, 3549, 97, 3581, 29954, 41588, 7422, 222, 31106, 48798, 6611, 6847, 226, 18594, 110, 7017, 237, 9709, 17665, 26460, 1543, 51165], "temperature": 0, "avg_logprob": -0.08200719702334805, "compression_ratio": 1.138996138996139, "no_speech_prob": 2.7547306449826614e-11}, {"id": 20, "seek": 17500, "start": 191.0, "end": 196.0, "text": "開源生態持續升溫,模型與開源領域迎來多項新進展。", "tokens": [51165, 8949, 47402, 8244, 45248, 17694, 29583, 41670, 16084, 104, 11, 41908, 39823, 34127, 8949, 47402, 8313, 246, 16262, 253, 17699, 3763, 6392, 8313, 227, 12560, 18214, 43491, 1543, 51415], "temperature": 0, "avg_logprob": -0.08200719702334805, "compression_ratio": 1.138996138996139, "no_speech_prob": 2.7547306449826614e-11}, {"id": 21, "seek": 19600, "start": 196.0, "end": 206.0, "text": "Vercel數據顯示,開源模型Token份額兩月內從28.4%飙升至62%,首次超越B元模型的38%。", "tokens": [50365, 53, 2869, 338, 30622, 36841, 46605, 25696, 11, 8949, 47402, 41908, 39823, 51, 8406, 36266, 10178, 235, 16313, 6939, 28472, 21068, 11205, 13, 19, 4, 11808, 247, 41670, 20844, 28052, 8923, 25444, 9487, 19869, 25761, 33, 14812, 41908, 39823, 1546, 12625, 4, 1543, 50865], "temperature": 0, "avg_logprob": -0.1340391402854059, "compression_ratio": 1.1083916083916083, "no_speech_prob": 2.698669933354214e-11}, {"id": 22, "seek": 19600, "start": 206.0, "end": 214.0, "text": "這一增長發生在B元需求同步加速的背景下,並非萎縮所致,反映企業採用開源趨勢正加速。", "tokens": [50865, 32260, 24228, 252, 15353, 14637, 8244, 3581, 33, 14812, 32535, 32718, 13089, 31429, 9990, 31217, 1546, 46329, 50218, 4438, 11, 35934, 12107, 10674, 236, 21030, 106, 5966, 6784, 112, 11, 22138, 1431, 254, 7384, 223, 27119, 6900, 94, 9254, 8949, 47402, 9759, 101, 7978, 95, 15789, 9990, 31217, 1543, 51265], "temperature": 0, "avg_logprob": -0.1340391402854059, "compression_ratio": 1.1083916083916083, "no_speech_prob": 2.698669933354214e-11}, {"id": 23, "seek": 19600, "start": 214.0, "end": 221.0, "text": "Cloudflare開源內部AI操作系統 Cloudflare OS基於Workers構建,整合上下文與系統。", "tokens": [51265, 32787, 3423, 543, 8949, 47402, 28472, 13470, 48698, 13167, 235, 11914, 25368, 33725, 8061, 3423, 543, 12731, 26008, 19488, 28846, 433, 43362, 34157, 11, 27662, 14245, 5708, 4438, 17174, 34127, 25368, 33725, 1543, 51615], "temperature": 0, "avg_logprob": -0.1340391402854059, "compression_ratio": 1.1083916083916083, "no_speech_prob": 2.698669933354214e-11}, {"id": 24, "seek": 22100, "start": 221.0, "end": 233.0, "text": "通過Gatekeepers限制AI寫入Isolate隔離敏感數據,構建安全殺箱,解決Agent數據訪問風險,為企業AI應用提供更安全環境。", "tokens": [50365, 19550, 8816, 38, 473, 43153, 43446, 25491, 48698, 4510, 104, 14028, 6802, 48278, 165, 1206, 32662, 7017, 237, 9709, 30622, 36841, 11, 43362, 34157, 45466, 45528, 11249, 109, 11, 17278, 33540, 32, 6930, 30622, 36841, 5396, 103, 11361, 22713, 49753, 11, 6344, 7384, 223, 27119, 48698, 20481, 9254, 20949, 3254, 249, 19002, 45466, 40063, 108, 47885, 1543, 50965], "temperature": 0, "avg_logprob": -0.09081391590397532, "compression_ratio": 1.138036809815951, "no_speech_prob": 2.5361324151873532e-11}, {"id": 25, "seek": 22100, "start": 233.0, "end": 244.0, "text": "開源AI Agent Hermes以YOLO模式全自動入侵泰國財政部,掃描漏洞並竊取2012年人士記錄,攻擊日誌含585個文件及憑證。", "tokens": [50965, 8949, 47402, 48698, 27174, 21842, 279, 3588, 56, 5046, 46, 41908, 27584, 11319, 9722, 12572, 14028, 3254, 113, 6847, 108, 8053, 11561, 94, 18523, 13470, 11, 6900, 225, 11673, 237, 14065, 237, 12136, 252, 35934, 11957, 232, 29436, 2009, 4762, 5157, 4035, 30337, 16958, 9567, 226, 11, 41530, 37813, 6890, 3549, 234, 2392, 104, 20, 19287, 3338, 17174, 20485, 25703, 29713, 239, 11823, 231, 1543, 51515], "temperature": 0, "avg_logprob": -0.09081391590397532, "compression_ratio": 1.138036809815951, "no_speech_prob": 2.5361324151873532e-11}, {"id": 26, "seek": 22100, "start": 244.0, "end": 250.0, "text": "事件凸顯無人監管Agent的自動化攻擊風險,引發責任歸屬爭議,", "tokens": [51515, 6973, 20485, 6336, 116, 46605, 16976, 4035, 5419, 96, 23131, 32, 6930, 1546, 9722, 12572, 23756, 41530, 37813, 22713, 49753, 11, 41301, 14637, 11561, 105, 26443, 4287, 116, 9636, 105, 8164, 255, 24686, 11, 51815], "temperature": 0, "avg_logprob": -0.09081391590397532, "compression_ratio": 1.138036809815951, "no_speech_prob": 2.5361324151873532e-11}, {"id": 27, "seek": 25000, "start": 250.0, "end": 253.0, "text": "涉及操作者、開發者與模型。", "tokens": [50365, 35681, 231, 25703, 13167, 235, 11914, 12444, 1231, 8949, 14637, 12444, 34127, 41908, 39823, 1543, 50515], "temperature": 0, "avg_logprob": -0.12539076979142905, "compression_ratio": 1.1278688524590164, "no_speech_prob": 3.388389915870427e-11}, {"id": 28, "seek": 25000, "start": 253.0, "end": 263.0, "text": "DeepMind推出Recirculation方法,無需重訓即可讓Transformer在推理時地規進化,通過預填充階段反饋激活實現動態狀態跟蹤。", "tokens": [50515, 11089, 595, 44, 471, 33597, 7781, 48883, 347, 2444, 399, 9249, 11148, 11, 16976, 32535, 12624, 5396, 241, 39127, 4429, 21195, 33339, 837, 260, 3581, 33597, 13876, 6611, 10928, 2888, 237, 18214, 23756, 11, 19550, 8816, 30156, 13331, 104, 2347, 227, 10673, 236, 28427, 22138, 41520, 233, 26373, 222, 25956, 10376, 9581, 12572, 45248, 18637, 222, 45248, 9678, 39540, 97, 1543, 51015], "temperature": 0, "avg_logprob": -0.12539076979142905, "compression_ratio": 1.1278688524590164, "no_speech_prob": 3.388389915870427e-11}, {"id": 29, "seek": 25000, "start": 263.0, "end": 269.0, "text": "生成成成本保持平坦,串行工作集中在預填充階段,突破層數限制。", "tokens": [51015, 8244, 11336, 11336, 11336, 8802, 24302, 17694, 16716, 14872, 99, 11, 940, 110, 8082, 41315, 26020, 5975, 3581, 30156, 13331, 104, 2347, 227, 10673, 236, 28427, 11, 40859, 39560, 9636, 97, 30622, 43446, 25491, 1543, 51315], "temperature": 0, "avg_logprob": -0.12539076979142905, "compression_ratio": 1.1278688524590164, "no_speech_prob": 3.388389915870427e-11}, {"id": 30, "seek": 25000, "start": 269.0, "end": 273.0, "text": "以上就是今天的全部內容了,感謝收看,我們明天見。", "tokens": [51315, 29497, 5620, 34947, 38714, 28472, 25750, 2289, 11, 40143, 34333, 11, 5884, 11100, 6135, 10470, 1543, 51515], "temperature": 0, "avg_logprob": -0.12539076979142905, "compression_ratio": 1.1278688524590164, "no_speech_prob": 3.388389915870427e-11}, {"id": 31, "seek": 27300, "start": 273.0, "end": 275.0, "text": "MING PAO CANADA // MING PAO TORONTO", "tokens": [50365, 44, 3017, 17718, 46, 22931, 45852, 29178, 376, 3017, 17718, 46, 314, 2483, 1928, 15427, 50465], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 32, "seek": 27300, "start": 275.0, "end": 276.0, "text": "MING PAO TORONTO", "tokens": [50465, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 50515], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 33, "seek": 27300, "start": 276.0, "end": 277.0, "text": "MING PAO TORONTO", "tokens": [50515, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 50565], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 34, "seek": 27300, "start": 277.0, "end": 278.0, "text": "MING PAO TORONTO", "tokens": [50565, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 50615], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 35, "seek": 27300, "start": 278.0, "end": 279.0, "text": "MING PAO TORONTO", "tokens": [50615, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 50665], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 36, "seek": 27300, "start": 279.0, "end": 280.0, "text": "MING PAO TORONTO", "tokens": [50665, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 50715], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 37, "seek": 27300, "start": 280.0, "end": 281.0, "text": "MING PAO TORONTO", "tokens": [50715, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 50765], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 38, "seek": 27300, "start": 281.0, "end": 282.0, "text": "MING PAO TORONTO", "tokens": [50765, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 50815], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 39, "seek": 27300, "start": 282.0, "end": 283.0, "text": "MING PAO TORONTO", "tokens": [50815, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 50865], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 40, "seek": 27300, "start": 283.0, "end": 284.0, "text": "MING PAO TORONTO", "tokens": [50865, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 50915], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 41, "seek": 27300, "start": 284.0, "end": 285.0, "text": "MING PAO TORONTO", "tokens": [50915, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 50965], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 42, "seek": 27300, "start": 285.0, "end": 286.0, "text": "MING PAO TORONTO", "tokens": [50965, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 51015], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 43, "seek": 27300, "start": 286.0, "end": 286.0, "text": "", "tokens": [], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11, "words": []}, {"id": 44, "seek": 27300, "start": 286.0, "end": 287.0, "text": "MING PAO TORONTO", "tokens": [51015, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 51065], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 45, "seek": 27300, "start": 287.0, "end": 288.0, "text": "MING PAO TORONTO", "tokens": [51065, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 51115], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 46, "seek": 27300, "start": 288.0, "end": 289.0, "text": "MING PAO TORONTO", "tokens": [51115, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 51165], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 47, "seek": 27300, "start": 289.0, "end": 290.0, "text": "MING PAO TORONTO", "tokens": [51165, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 51215], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 48, "seek": 27300, "start": 290.0, "end": 291.0, "text": "MING PAO TORONTO", "tokens": [51215, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 51265], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 49, "seek": 27300, "start": 291.0, "end": 292.0, "text": "MING PAO TORONTO", "tokens": [51265, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 51315], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 50, "seek": 27300, "start": 292.0, "end": 293.0, "text": "MING PAO TORONTO", "tokens": [51315, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 51365], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}, {"id": 51, "seek": 27300, "start": 293.0, "end": 294.0, "text": "MING PAO TORONTO", "tokens": [51365, 44, 3017, 17718, 46, 314, 2483, 1928, 15427, 51415], "temperature": 0, "avg_logprob": -0.0747382843975529, "compression_ratio": 8.853658536585366, "no_speech_prob": 2.9031159420878083e-11}], "language": "zh"}