{"id":40098,"date":"2025-04-13T01:02:45","date_gmt":"2025-04-12T17:02:45","guid":{"rendered":"https:\/\/17aitech.com\/?p=40098"},"modified":"2025-04-13T01:02:45","modified_gmt":"2025-04-12T17:02:45","slug":"%e8%8b%b9%e6%9e%9c%e5%8f%91%e7%8e%b0%e5%8e%9f%e7%94%9f%e5%a4%9a%e6%a8%a1%e6%80%81%e6%a8%a1%e5%9e%8bscaling-laws%ef%bc%9a%e6%97%a9%e8%9e%8d%e5%90%88%e4%bc%98%e4%ba%8e%e5%90%8e%e8%9e%8d%e5%90%88","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=40098","title":{"rendered":"\u82f9\u679c\u53d1\u73b0\u539f\u751f\u591a\u6a21\u6001\u6a21\u578bScaling Laws\uff1a\u65e9\u878d\u5408\u4f18\u4e8e\u540e\u878d\u5408\uff0cMoE\u4f18\u4e8e\u5bc6\u96c6\u6a21\u578b"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-04-12-3\" target=\"_blank\">\u82f9\u679c\u53d1\u73b0\u539f\u751f\u591a\u6a21\u6001\u6a21\u578bScaling Laws\uff1a\u65e9\u878d\u5408\u4f18\u4e8e\u540e\u878d\u5408\uff0cMoE\u4f18\u4e8e\u5bc6\u96c6\u6a21\u578b<\/a><\/p>\n<p>\u8ba9\u5927\u6a21\u578b\u8fdb\u5165\u591a\u6a21\u6001\u6a21\u5f0f\uff0c\u4ece\u800c\u80fd\u591f\u6709\u6548\u611f\u77e5\u4e16\u754c\uff0c\u662f\u6700\u8fd1 AI \u9886\u57df\u91cc\u4eba\u4eec\u4e00\u76f4\u7684\u63a2\u7d22\u76ee\u6807\u3002<\/p>\n<p>\u76ee\u524d\u6211\u4eec\u89c1\u5230\u7684\u5f88\u591a\u591a\u6a21\u6001\u5927\u6a21\u578b\u5e94\u7528\u662f\u300c\u7ec4\u5408\u5f0f\u300d\u7684\uff1a\u5176\u4e2d\u96c6\u6210\u4e86\u6570\u4e2a\u5355\u72ec\u9884\u8bad\u7ec3\u7684\u7ec4\u4ef6\uff0c\u4f8b\u5982\u5c06\u89c6\u89c9\u7f16\u7801\u5668\u8fde\u63a5\u5230 LLM \u4e0a\u5e76\u7ee7\u7eed\u8fdb\u884c\u591a\u6a21\u6001\u8bad\u7ec3\uff1b\u800c\u5728\u8c37\u6b4c Gemin 2.0 \u63a8\u51fa\u4e4b\u540e\uff0c\u539f\u751f\u591a\u6a21\u6001\u6a21\u578b\uff08NMM\uff09\u88ab\u8ba4\u4e3a\u662f\u6b63\u786e\u7684\u65b9\u5411\u3002<\/p>\n<p>\u4f46\u4ece\u96f6\u5f00\u59cb\u8bad\u7ec3\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u771f\u7684\u662f\u66f4\u597d\u7684\u65b9\u6cd5\u5417\uff1f\u8fd1\u65e5\uff0c\u6765\u81ea\u6cd5\u56fd\u7d22\u90a6\u5927\u5b66\u3001\u82f9\u679c\u7684\u7814\u7a76\u8005\u8fdb\u884c\u4e86\u4e00\u9879\u5e7f\u6cdb\u7684 Scaling Laws \u7814\u7a76\uff0c\u6db5\u76d6\u4e86 457 \u4e2a\u91c7\u7528\u4e0d\u540c\u67b6\u6784\u548c\u8bad\u7ec3\u65b9\u5f0f\u7684\u6a21\u578b\u3002<\/p>\n<p>\u7814\u7a76\u8868\u660e\uff0c\u4e0d\u4f9d\u8d56\u56fe\u50cf\u7f16\u7801\u5668\u7684\u65e9\u878d\u5408\u67b6\u6784\u4e0e\u540e\u878d\u5408\u67b6\u6784\u76f8\u6bd4\u5e76\u6ca1\u6709\u56fa\u6709\u4f18\u52bf\u3002\u4e0d\u8fc7\uff0c<strong>\u65e9\u878d\u5408\u5728\u8f83\u4f4e\u7684\u53c2\u6570\u6570\u91cf\u4e0b\u8868\u73b0\u51fa\u66f4\u5f3a\u7684\u6027\u80fd\uff0c\u8bad\u7ec3\u6548\u7387\u66f4\u9ad8\uff0c\u5e76\u4e14\u66f4\u6613\u4e8e\u90e8\u7f72\u3002<\/strong>\u53d7\u65e9\u878d\u5408\u67b6\u6784\u5f3a\u5927\u6027\u80fd\u7684\u542f\u53d1\uff0c\u5b9e\u9a8c\u8868\u660e\uff0c<strong>\u7ed3\u5408\u6df7\u5408\u4e13\u5bb6 (MoE) \u53ef\u4ee5\u4f7f\u6a21\u578b\u5b66\u4e60\u7279\u5b9a\u4e8e\u6a21\u6001\u7684\u6743\u91cd\uff0c\u4ece\u800c\u663e\u8457\u63d0\u5347\u6027\u80fd\u3002<\/strong><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dd8da943491889b92c64420bfe784029.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dd8da943491889b92c64420bfe784029.png\"><\/a><\/section>\n<ul>\n<li>\n<p>\u8bba\u6587\u6807\u9898\uff1aScaling Laws for Native Multimodal Models<\/p>\n<\/li>\n<li>\n<p>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2504.07951<\/p>\n<\/li>\n<\/ul>\n<p>\u7814\u7a76\u4eba\u5458\u7684\u53d1\u73b0\u53ef\u4ee5\u6982\u62ec\u5982\u4e0b\uff1a<strong>\u539f\u751f\u65e9\u878d\u5408\u548c\u540e\u878d\u5408\u6a21\u578b\u7684\u6027\u80fd\u76f8\u5f53<\/strong>\uff0c\u4ece\u96f6\u5f00\u59cb\u8bad\u7ec3\u7684\u65e9\u878d\u5408\u6a21\u578b\u6027\u80fd\u4e0e\u540e\u878d\u5408\u6a21\u578b\u6027\u80fd\u76f8\u5f53\uff0c\u5728\u8ba1\u7b97\u9884\u7b97\u8f83\u4f4e\u7684\u60c5\u51b5\u4e0b\uff0c\u7565\u4f18\u4e8e\u65e9\u878d\u5408\u6a21\u578b\u3002\u6b64\u5916\uff0c\u5bf9\u4e8e\u591a\u6a21\u6001 Scaling Laws \u7684\u7814\u7a76\u8868\u660e\uff0c\u968f\u7740\u8ba1\u7b97\u9884\u7b97\u7684\u589e\u52a0\uff0c\u65e9\u878d\u5408\u548c\u540e\u878d\u5408\u7684\u8ba1\u7b97\u6700\u4f18\u6a21\u578b\u6027\u80fd\u76f8\u4f3c\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-a064ad2d86e62401f29719c9ab94535c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-a064ad2d86e62401f29719c9ab94535c.png\"><\/a><\/section>\n<p><strong>NMM \u7684 Scaling Law \u4e0e LLM \u7c7b\u4f3c<\/strong>\uff1a\u539f\u751f\u591a\u6a21\u6001\u6a21\u578b\u7684 Scaling Law \u9075\u5faa\u4e0e\u7eaf\u6587\u672c LLM \u7c7b\u4f3c\u7684\u89c4\u5f8b\uff0cscaling \u6307\u6570\u6839\u636e\u76ee\u6807\u6570\u636e\u7c7b\u578b\u548c\u8bad\u7ec3\u7ec4\u5408\u7565\u6709\u4e0d\u540c\u3002<\/p>\n<p>\u4e0e\u6b64\u540c\u65f6\uff0c<strong>\u540e\u878d\u5408\u7684\u6a21\u578b\u9700\u8981\u66f4\u591a\u53c2\u6570<\/strong>\uff1a\u4e0e\u65e9\u878d\u5408\u76f8\u6bd4\uff0c\u8ba1\u7b97\u6700\u4f18\u7684\u540e\u878d\u5408\u6a21\u578b\u9700\u8981\u66f4\u9ad8\u7684\u53c2\u6570\u4e0e\u6570\u636e\u6bd4\uff08\u89c1\u56fe 1 \u53f3\uff09\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-79364b1eb2cf2a321d3ab0079c38ce67.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-79364b1eb2cf2a321d3ab0079c38ce67.png\"><\/a><\/section>\n<p>\u518d\u5f80\u4e0b\uff0c<strong>\u7a00\u758f\u6027\u663e\u8457\u6709\u5229\u4e8e\u65e9\u878d\u5408\u7684 NMM<\/strong>\uff1a\u5728\u76f8\u540c\u7684\u63a8\u7406\u6210\u672c\u4e0b\uff0c\u7a00\u758f NMM \u4e0e\u5bc6\u96c6 NMM \u76f8\u6bd4\u8868\u73b0\u51fa\u663e\u8457\u7684\u6539\u8fdb\u3002<\/p>\n<p>\u5728\u7a00\u758f\u6027\u8bad\u7ec3\u4e2d\uff0c\u5b83\u4eec\u4f1a\u9690\u5f0f\u5b66\u4e60\u7279\u5b9a\u4e8e\u6a21\u6001\u7684\u6743\u91cd\uff08\u56fe 23\uff09\u3002\u6b64\u5916\uff0c\u968f\u7740\u8ba1\u7b97\u9884\u7b97\u7684\u589e\u957f\uff0c\u8ba1\u7b97\u4f18\u5316\u6a21\u578b\u66f4\u591a\u5730\u4f9d\u8d56\u4e8e\u8bad\u7ec3 token \u6570\u91cf\u7684\u6269\u5c55\uff0c\u800c\u4e0d\u662f\u6d3b\u52a8\u53c2\u6570\u7684\u6570\u91cf\uff08\u56fe 1 \u53f3\uff09\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-1d66283c0e4dff5ae7f3720a8dc6aa78.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-1d66283c0e4dff5ae7f3720a8dc6aa78.png\"><\/a><\/section>\n<p><strong>\u5bf9\u4e8e\u7a00\u758f NMM\uff0c\u6a21\u6001\u65e0\u5173\u8def\u7531\u4f18\u4e8e\u6a21\u6001\u611f\u77e5\u8def\u7531<\/strong>\uff1a\u4f7f\u7528\u6a21\u6001\u65e0\u5173\u8def\u7531\u8bad\u7ec3\u7a00\u758f\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff0c\u5176\u6027\u80fd\u59cb\u7ec8\u4f18\u4e8e\u4f7f\u7528\u6a21\u6001\u611f\u77e5\u8def\u7531\u7684\u6a21\u578b\uff08\u56fe 11\uff09\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-37626215be83545ec092757239c73e93.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-37626215be83545ec092757239c73e93.png\"><\/a><\/section>\n<p><strong>\u539f\u751f\u591a\u6a21\u6001\u6a21\u578b\u7684 scaling \u5c5e\u6027<\/strong><\/p>\n<p>\u4e0b\u9762\u5177\u4f53\u5c55\u793a\u539f\u751f\u591a\u6a21\u6001\u6a21\u578b\u7684 Scaling Laws\u3002\u8be5\u56e2\u961f\u7814\u7a76\u4e86\u591a\u79cd\u67b6\u6784\u9009\u62e9\u3001\u4e0d\u540c\u7684\u6570\u636e\u6df7\u5408\u65b9\u5f0f\u3001\u65e9\u878d\u5408\u548c\u540e\u878d\u5408 NMM \u4e4b\u95f4\u7684\u5b9e\u9645\u6743\u8861\u3001NMM \u7684\u539f\u751f\u9884\u8bad\u7ec3\u548c\u6301\u7eed\u9884\u8bad\u7ec3\u7684\u6027\u80fd\u8868\u73b0\u3002<\/p>\n<p><strong>NMM \u7684 Scaling Laws<\/strong><\/p>\n<p>\u65e9\u878d\u5408\u548c\u540e\u878d\u5408\u6a21\u578b\u7684 Scaling Laws\u3002<\/p>\n<p>\u56fe 2 \u5de6\u56fe\u5c55\u793a\u4e86\u65e9\u878d\u5408 NMM \u5728\u591a\u6a21\u6001\u4ea4\u7ec7\u6570\u636e\u96c6\u3001\u56fe\u50cf &#8211; \u63cf\u8ff0\u6570\u636e\u96c6\u548c\u6587\u672c\u6570\u636e\u96c6\u4e0a\u7684\u5e73\u5747\u6700\u7ec8\u635f\u5931\u3002\u5176\u6700\u4f4e\u635f\u5931\u8fb9\u754c\u9075\u5faa\u4e0e FLOPs \u7684\u5e42\u5f8b\u5173\u7cfb\u3002\u62df\u5408\u8be5\u5e42\u5f8b\u53ef\u5f97\u51fa\u8868\u8fbe\u5f0f L \u221d C^\u22120.049\uff0c\u8fd9\u6307\u793a\u4e86\u968f\u7740\u8ba1\u7b97\u91cf\u7684\u589e\u52a0\uff0c\u6027\u80fd\u63d0\u5347\u7684\u901f\u7387\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dd5f2b9288be0a287e1a8f92f55c9481.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dd5f2b9288be0a287e1a8f92f55c9481.png\"><\/a><\/section>\n<p>\u62df\u5408\u8be5\u5e42\u5f8b\u53ef\u5f97\u51fa\u8868\u8fbe\u5f0f L \u221d C^\u22120.049\uff0c\u8fd9\u6307\u793a\u4e86\u968f\u7740\u8ba1\u7b97\u91cf\u7684\u589e\u52a0\uff0c\u6027\u80fd\u63d0\u5347\u7684\u901f\u7387\u3002<\/p>\n<p>\u5206\u6790\u6bcf\u79cd\u6570\u636e\u7c7b\u578b\uff08\u4f8b\u5982\uff0c\u56fe\u50cf &#8211; \u8bf4\u660e\u3001\u591a\u6a21\u6001\u4ea4\u7ec7\u7684\u6587\u6863\u3001\u6587\u672c\uff09\u7684 Scaling Laws \u65f6\uff0c\u53ef\u4ee5\u89c2\u5bdf\u5230\u6307\u6570\u4f1a\u53d1\u751f\u53d8\u5316\uff08\u8868 3\uff09\u3002\u4f8b\u5982\uff0c\u4e0e\u591a\u6a21\u6001\u4ea4\u7ec7\uff08L \u221d C^\u22120.046\uff09\u76f8\u6bd4\uff0c\u8be5\u6a21\u578b\u5728\u56fe\u50cf\u8bf4\u660e\u6570\u636e\uff08L \u221d C^\u22120.061\uff09\u4e0a\u5b9e\u73b0\u4e86\u66f4\u9ad8\u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-850fca8e87716aa0066ca8e2f1153afc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-850fca8e87716aa0066ca8e2f1153afc.png\"><\/a><\/section>\n<p>\u4e3a\u4e86\u5c06\u635f\u5931\u5efa\u6a21\u4e3a\u8bad\u7ec3 token \u6570\u91cf D \u548c\u6a21\u578b\u53c2\u6570 N \u7684\u51fd\u6570\uff0c\u4ed6\u4eec\u8fd8\u62df\u5408\u4e86\u4ee5\u4e0b\u53c2\u6570\u51fd\u6570\uff0c\u5f97\u5230\u4e86 scaling \u6307\u6570 \u03b1 = 0.301 \u548c \u03b2 = 0.335\u3002\u5b83\u4eec\u5206\u522b\u63cf\u8ff0\u4e86 scaling \u6a21\u578b\u53c2\u6570\u548c\u8bad\u7ec3 token \u6570\u91cf\u65f6\u7684\u6027\u80fd\u63d0\u5347\u7387\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-b3a4d18f94904d5e2f9b2e6e0a6098b5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-b3a4d18f94904d5e2f9b2e6e0a6098b5.png\"><\/a><\/section>\n<p>\u8be5\u56e2\u961f\u5047\u8bbe\u8ba1\u7b97\u91cf\u3001N \u548c D \u4e4b\u95f4\u5b58\u5728\u7ebf\u6027\u5173\u7cfb\uff08\u5373 C \u221d N D\uff09\uff0c\u63a8\u5bfc\u51fa\u4e86\u6a21\u578b\u53c2\u6570\u4e0e\u8ba1\u7b97\u9884\u7b97\u4e4b\u95f4\u7684\u76f8\u5173\u89c4\u5f8b\uff0c\u8be6\u89c1\u539f\u8bba\u6587\u9644\u5f55 C\u3002\u7b80\u5355\u6765\u8bf4\uff0c\u5bf9\u4e8e\u7ed9\u5b9a\u7684\u8ba1\u7b97\u9884\u7b97 C\uff0c\u4ee5\u5bf9\u6570\u95f4\u9694\u7684 D \u503c\u8ba1\u7b97\u76f8\u5e94\u7684\u6a21\u578b\u5927\u5c0f N\uff0c\u5e76\u786e\u5b9a\u4f7f\u635f\u5931\u6700\u5c0f\u5316\u7684\u53c2\u6570\u6570\u91cf N_opt\u3002\u5bf9\u4e0d\u540c\u7684 FLOPs \u503c\u91cd\u590d\u6b64\u64cd\u4f5c\uff0c\u53ef\u5f97\u5230\u4e00\u4e2a (C, N_opt) \u6570\u636e\u96c6\uff0c\u8be5\u56e2\u961f\u5bf9\u5176\u62df\u5408\u4e86\u4e00\u4e2a\u5e42\u5f8b\uff0c\u53ef\u9884\u6d4b\u8ba1\u7b97\u6700\u4f18\u6a21\u578b\u5927\u5c0f\u4e0e\u8ba1\u7b97\u91cf\u7684\u5173\u7cfb\uff1aN \u2217 \u221d C^0.526\u3002<\/p>\n<p>\u7c7b\u4f3c\u5730\uff0c\u4ed6\u4eec\u4e5f\u62df\u5408\u4e86\u4f30\u8ba1\u8ba1\u7b97\u6700\u4f18\u8bad\u7ec3\u6570\u636e\u96c6\u5927\u5c0f\u4e0e\u8ba1\u7b97\u91cf\u548c\u6a21\u578b\u5927\u5c0f\u7684\u5173\u7cfb\uff1a<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-7a69137c6661f22f98e3cb66f2edc996.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-7a69137c6661f22f98e3cb66f2edc996.png\"><\/a><\/section>\n<p>\u8fd9\u4e9b\u5173\u7cfb\u53ef\u8ba9\u5b9e\u8df5\u8005\u5728\u7ed9\u5b9a\u56fa\u5b9a\u8ba1\u7b97\u9884\u7b97\u7684\u60c5\u51b5\u4e0b\u786e\u5b9a\u6700\u4f73\u6a21\u578b\u548c\u6570\u636e\u96c6\u5927\u5c0f\u3002\u5728\u6839\u636e\u6570\u636e\u7c7b\u578b\u5206\u6790\u65f6\uff0c\u8be5\u56e2\u961f\u53d1\u73b0\u4e0e\u56fe\u50cf\u8bf4\u660e\u6570\u636e (a = 0.520) \u76f8\u6bd4\uff0c\u591a\u6a21\u6001\u4ea4\u7ec7\u6570\u636e\u53ef\u4ece\u8f83\u5927\u7684\u6a21\u578b (a = 0.532) \u4e2d\u83b7\u76ca\u66f4\u591a\uff0c\u800c\u8bad\u7ec3 token \u5219\u5448\u73b0\u76f8\u53cd\u7684\u8d8b\u52bf\u3002<\/p>\n<p>\u8be5\u56e2\u961f\u5728\u56fe 2\uff08\u53f3\uff09\u4e2d\u5bf9\u540e\u878d\u5408\u6a21\u578b\u8fdb\u884c\u4e86\u7c7b\u4f3c\u7684\u7814\u7a76\uff0c\u5e76\u89c2\u5bdf\u5230\u4e86\u7c7b\u4f3c\u7684 scaling \u884c\u4e3a\u3002\u5177\u4f53\u800c\u8a00\uff0c\u635f\u5931 scaling \u6307\u6570 (c = \u22120.0494) \u4e0e\u65e9\u878d\u5408 (c = \u22120.0492) \u51e0\u4e4e\u76f8\u540c\u3002\u8fd9\u4e00\u8d8b\u52bf\u5728\u56fe 3 \u4e2d\u663e\u800c\u6613\u89c1\uff0c\u5728\u8f83\u5c0f\u7684\u6a21\u578b\u89c4\u6a21\u4e0b\uff0c\u65e9\u878d\u5408\u7684\u8868\u73b0\u4f18\u4e8e\u540e\u878d\u5408\uff0c\u800c\u5728\u8f83\u5927\u7684\u6a21\u578b\u89c4\u6a21\u4e0b\uff0c\u4e24\u79cd\u67b6\u6784\u7684\u6027\u80fd\u6536\u655b\u5230\u76f8\u4f3c\u7684\u6c34\u5e73\u3002\u5728\u6539\u53d8\u540e\u878d\u5408\u914d\u7f6e\u65f6\uff0c\u8be5\u56e2\u961f\u4e5f\u89c2\u5bdf\u5230\u4e86\u7c7b\u4f3c\u7684\u8d8b\u52bf\uff0c\u4f8b\u5982\u4f7f\u7528\u8f83\u5c0f\u7684\u89c6\u89c9\u7f16\u7801\u5668\u548c\u8f83\u5927\u7684\u6587\u672c\u89e3\u7801\u5668\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-2c6d68616e25b679162df886d622a1bd.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-2c6d68616e25b679162df886d622a1bd.png\"><\/a><\/section>\n<p>NMM \u548c LLM \u7684 Scaling Laws \u6bd4\u8f83\u3002<\/p>\n<p>\u901a\u8fc7\u6bd4\u8f83 NMM \u7684 scaling laws \u7cfb\u6570\u4e0e\u7eaf\u6587\u672c LLM\uff08\u4f8b\u5982 GPT-3\u3001Chinchilla\uff09\u7684 scaling laws \u7cfb\u6570\uff0c\u8be5\u56e2\u961f\u53d1\u73b0\u5b83\u4eec\u5904\u4e8e\u76f8\u4f3c\u7684\u8303\u56f4\u5185\u3002<\/p>\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u5982\u679c\u5c06\u635f\u5931\u770b\u4f5c\u8ba1\u7b97\u91cf\u7684\u51fd\u6570\uff0cGPT-3 \u9075\u5faa L \u221d C^\u22120.048\uff0c\u800c\u8be5\u56e2\u961f\u7684\u6a21\u578b\u9075\u5faa L \u221d C^\u22120.049\uff0c\u8fd9\u8868\u660e NMM \u7684\u6027\u80fd\u9075\u5faa\u4e0e LLM \u7c7b\u4f3c\u7684 scaling \u89c4\u5f8b\u3002\u540c\u6837\uff0c\u8be5\u56e2\u961f\u5bf9 \u03b1 \u548c \u03b2 \u53c2\u6570\u7684\u4f30\u8ba1\u503c (\u03b1 = 0.301, \u03b2 = 0.335) \u4e0e Hoffmann \u7b49\u4eba\u62a5\u544a\u7684\u503c (\u03b1 = 0.339, \u03b2 = 0.285) \u975e\u5e38\u63a5\u8fd1\u3002\u540c\u6837\uff0c\u8be5\u56e2\u961f\u8ba1\u7b97\u51fa\u7684 a = 0.526 \u548c b = 0.473 \u4e0e Hoffmann \u7b49\u4eba\u7684 a = 0.46 \u548c b = 0.54 \u975e\u5e38\u63a5\u8fd1\u3002<\/p>\n<p>\u8fd9\u8868\u660e\uff1a\u5bf9\u4e8e\u539f\u751f\u591a\u6a21\u6001\u6a21\u578b\uff0c\u8bad\u7ec3 token \u7684\u6570\u91cf\u548c\u6a21\u578b\u53c2\u6570\u5e94\u6309\u6bd4\u4f8b scaling\u3002\u7136\u800c\uff0c\u7531\u4e8e a \u548c b \u4e4b\u95f4\u7684\u5dee\u8ddd\u5c0f\u4e8e LLM\uff0c\u56e0\u6b64\u8fd9\u4e00\u539f\u5219\u5bf9\u4e8e NMM \u66f4\u4e3a\u9002\u7528\u3002\u6b64\u5916\uff0c\u7531\u4e8e\u5728\u8be5\u56e2\u961f\u7684\u6848\u4f8b\u4e2d a = 0.526 \u5927\u4e8e b = 0.473\uff0c\u56e0\u6b64\u5728\u8ba1\u7b97\u9884\u7b97\u56fa\u5b9a\u7684\u60c5\u51b5\u4e0b\uff0c<strong>NMM \u7684\u6700\u4f73\u6a21\u578b\u5927\u5c0f\u5927\u4e8e LLM\uff0c\u800c\u6700\u4f73\u8bad\u7ec3 token \u6570\u91cf\u5219\u8f83\u4f4e\u3002<\/strong><\/p>\n<p>\u65e9\u878d\u5408\u4e0e\u540e\u878d\u5408 NMM \u7684\u8ba1\u7b97\u4f18\u5316\u6743\u8861\u3002<\/p>\n<p>\u867d\u7136\u540e\u878d\u5408\u548c\u65e9\u878d\u5408\u6a21\u578b\u7684\u635f\u5931\u4f1a\u968f\u7740 FLOP \u7684\u589e\u52a0\u4ee5\u76f8\u4f3c\u7684\u901f\u7387\u964d\u4f4e\uff0c\u4f46\u8be5\u56e2\u961f\u89c2\u5bdf\u5230\u5b83\u4eec\u7684\u8ba1\u7b97\u4f18\u5316\u6a21\u578b\u4e2d\u5b58\u5728\u660e\u663e\u7684\u6743\u8861\u3002\u5177\u4f53\u800c\u8a00\uff0c\u540e\u878d\u5408\u6a21\u578b\u7684 N_opt \u8f83\u5927\uff0c\u800c\u65e9\u878d\u5408\u6a21\u578b\u7684 D_opt \u8f83\u5927\u3002\u8fd9\u8868\u660e\uff0c\u5728\u8ba1\u7b97\u9884\u7b97\u56fa\u5b9a\u7684\u60c5\u51b5\u4e0b\uff0c\u540e\u878d\u5408\u6a21\u578b\u9700\u8981\u66f4\u591a\u53c2\u6570\uff0c\u800c\u65e9\u878d\u5408\u6a21\u578b\u5219\u53d7\u76ca\u4e8e\u66f4\u591a\u8bad\u7ec3 token\u3002<\/p>\n<p>\u8fd9\u4e00\u8d8b\u52bf\u4e5f\u53cd\u6620\u5728\u65e9\u878d\u5408\u7684 <a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-2894e9901f9382b3dec42a9e9589c6ac.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-2894e9901f9382b3dec42a9e9589c6ac.png\"><\/a> \u4f4e\u4e8e\u540e\u878d\u5408\u7684 <a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-e51043826035f6f72b10b4952a8ccd5b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-e51043826035f6f72b10b4952a8ccd5b.png\"><\/a>\u00a0\u3002\u5982\u56fe 1\uff08\u53f3\uff09\u6240\u793a\uff0c\u968f\u7740 FLOPs \u7684 scaling\uff0c\u65e9\u878d\u5408\u6a21\u578b\u7684\u53c2\u6570\u6570\u91cf\u663e\u8457\u51cf\u5c11\uff0c\u8fd9\u5bf9\u4e8e\u964d\u4f4e\u63a8\u7406\u6210\u672c\uff0c\u4ece\u800c\u964d\u4f4e\u90e8\u7f72\u540e\u7684\u670d\u52a1\u6210\u672c\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p>\u65e9\u878d\u5408\u7684\u8bad\u7ec3\u6548\u7387\u66f4\u9ad8\u3002<\/p>\n<p>\u8be5\u56e2\u961f\u6bd4\u8f83\u4e86\u540e\u878d\u5408\u548c\u65e9\u878d\u5408\u67b6\u6784\u7684\u8bad\u7ec3\u6548\u7387\u3002\u5982\u56fe 5 \u6240\u793a\uff0c\u5728\u76f8\u540c\u7684\u8ba1\u7b97\u9884\u7b97\u4e0b\uff0c\u65e9\u878d\u5408\u6a21\u578b\u6d88\u8017\u7684\u5185\u5b58\u66f4\u5c11\uff0c\u8bad\u7ec3\u901f\u5ea6\u66f4\u5feb\u3002\u968f\u7740\u8ba1\u7b97\u91cf\u7684\u589e\u52a0\uff0c\u8fd9\u4e00\u4f18\u52bf\u66f4\u52a0\u660e\u663e\uff0c\u8fd9\u8868\u660e\uff1a\u65e9\u878d\u5408\u5728\u4fdd\u6301\u4e0e\u540e\u878d\u5408\u76f8\u5f53\u7684\u5927\u89c4\u6a21\u6027\u80fd\u7684\u540c\u65f6\uff0c\u62e5\u6709\u5353\u8d8a\u7684\u8bad\u7ec3\u6548\u7387\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-3acb042223861dfb25060deedd0ac45e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-3acb042223861dfb25060deedd0ac45e.png\"><\/a><\/section>\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5728\u76f8\u540c\u7684 FLOPs \u4e0b\uff0c\u4e0e\u65e9\u878d\u5408\u6a21\u578b\u76f8\u6bd4\uff0c\u540e\u878d\u5408\u6a21\u578b\u5177\u6709\u66f4\u9ad8\u7684\u53c2\u6570\u6570\u91cf\u548c\u66f4\u9ad8\u7684\u6709\u6548\u6df1\u5ea6\uff08\u5373\u9664\u4e86\u89e3\u7801\u5668\u5c42\u4e4b\u5916\u8fd8\u589e\u52a0\u4e86\u989d\u5916\u7684\u89c6\u89c9\u7f16\u7801\u5668\u5c42\uff09\u3002<\/p>\n<p><strong>\u4e0d\u540c\u6570\u636e\u6df7\u5408\u65b9\u5f0f\u7684 Scaling Laws<\/strong><\/p>\n<p>\u56fe 4 \u5c55\u793a\u4e86\u4e0d\u540c\u7684\u6df7\u5408\u65b9\u5f0f\u90fd\u9075\u5faa\u76f8\u4f3c\u7684 scaling \u8d8b\u52bf\uff1b\u7136\u800c\uff0cscaling \u7cfb\u6570\u4f1a\u6709\u5dee\u522b\uff08\u8868 4\uff09\u3002\u6709\u8da3\u7684\u662f\uff0c\u589e\u52a0\u56fe\u50cf &#8211; \u8bf4\u660e\u6570\u636e\u7684\u6bd4\u4f8b\uff08mixtures 1 \u548c 2\uff09\u4f1a\u5bfc\u81f4 a \u964d\u4f4e\u3001b \u5347\u9ad8\uff0c\u800c\u589e\u52a0\u591a\u6a21\u6001\u4ea4\u7ec7\u6570\u636e\u548c\u6587\u672c\u6570\u636e\u7684\u6bd4\u4f8b\uff08mixtures 3 \u548c 4\uff09\u5219\u4f1a\u4ea7\u751f\u76f8\u53cd\u7684\u6548\u679c\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-615b9a855fd7ebae5526e28423135aef.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-615b9a855fd7ebae5526e28423135aef.png\"><\/a><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dcb96b62a7d50cff67efac9174f797b4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dcb96b62a7d50cff67efac9174f797b4.png\"><\/a><\/section>\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u56fe\u50cf\u8bf4\u660e\u6570\u636e\u5305\u542b\u7684\u56fe\u50cf token \u6bd4\u6587\u672c token \u591a\uff1b\u56e0\u6b64\uff0c\u589e\u52a0\u5176\u6bd4\u4f8b\u4f1a\u5bfc\u81f4\u56fe\u50cf token \u589e\u591a\uff0c\u800c\u589e\u52a0\u591a\u6a21\u6001\u4ea4\u7ec7\u6570\u636e\u548c\u6587\u672c\u6570\u636e\u7684\u6bd4\u4f8b\u4f1a\u589e\u52a0\u6587\u672c token \u7684\u6570\u91cf\u3002\u8fd9\u8868\u660e\uff0c\u5f53\u56fe\u50cf token \u5360\u4e3b\u5bfc\u5730\u4f4d\u65f6\uff0c\u8bad\u7ec3\u65f6\u95f4\u8d8a\u957f\uff0c\u635f\u5931\u7684\u964d\u4f4e\u901f\u5ea6\u5c31\u8d8a\u5feb\uff0c\u800c\u589e\u52a0\u6a21\u578b\u5927\u5c0f\u5219\u66f4\u5feb\u3002<\/p>\n<p>\u8be5\u56e2\u961f\u8fd8\u53d1\u73b0\uff0c\u5bf9\u4e8e\u56fa\u5b9a\u7684\u6a21\u578b\u5927\u5c0f\uff0c\u589e\u52a0\u7eaf\u6587\u672c\u548c\u591a\u6a21\u6001\u4ea4\u7ec7\u6570\u636e\u7684\u6bd4\u4f8b\u6709\u5229\u4e8e\u65e9\u878d\u5408\uff08\u56fe 6\uff09\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-42b5256b7b5f7fc10e68528f0d3675a4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-42b5256b7b5f7fc10e68528f0d3675a4.png\"><\/a><\/section>\n<p><strong>\u539f\u751f\u591a\u6a21\u6001\u9884\u8bad\u7ec3 vs. LLM \u7684\u6301\u7eed\u8bad\u7ec3<\/strong><\/p>\n<p>\u4e0b\u9762\u6bd4\u8f83\u8fd9\u4e24\u79cd\u60c5\u51b5\uff1a\u4ece\u96f6\u5f00\u59cb\u8fdb\u884c\u539f\u751f\u8bad\u7ec3\uff0c\u4ee5\u53ca\u4f7f\u7528\u9884\u8bad\u7ec3\u7684 LLM \u8fdb\u884c\u521d\u59cb\u5316\u540e\u518d\u8fdb\u884c\u6301\u7eed\u8bad\u7ec3\u3002<\/p>\n<p>\u8fd9\u91cc\u4f7f\u7528\u7684\u521d\u59cb\u6a21\u578b\u662f DCLM-1B\uff0c\u8be5\u6a21\u578b\u5df2\u4f7f\u7528\u8d85\u8fc7 2T \u4e2a token \u8fdb\u884c\u8bad\u7ec3\u3002\u56fe 8 \u8868\u660e\uff0c\u5f53\u8bad\u7ec3\u65f6\u95f4\u66f4\u957f\u65f6\uff0c\u539f\u751f\u591a\u6a21\u6001\u6a21\u578b\u53ef\u4ee5\u7f29\u5c0f\u4e0e\u521d\u59cb\u5316\u6a21\u578b\u7684\u5dee\u8ddd\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-5c99639ea0ca749d97a240b9fb5728be.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-5c99639ea0ca749d97a240b9fb5728be.png\"><\/a><\/section>\n<p>\u5177\u4f53\u800c\u8a00\uff0c\u5728\u56fe\u50cf\u8bf4\u660e\u6570\u636e\u4e0a\uff0c\u8be5\u6a21\u578b\u9700\u8981\u4e0d\u5230 100B \u4e2a\u591a\u6a21\u6001 token \u5373\u53ef\u8fbe\u5230\u76f8\u5f53\u7684\u6027\u80fd\u3002\u7136\u800c\uff0c\u5728\u591a\u6a21\u6001\u4ea4\u7ec7\u6570\u636e\u548c\u6587\u672c\u6570\u636e\u4e0a\uff0c\u8be5\u6a21\u578b\u53ef\u80fd\u9700\u8981\u66f4\u957f\u65f6\u95f4\u7684\u8bad\u7ec3 \u2014\u2014 \u53ef\u591a\u8fbe 1T \u4e2a token\u3002\u8003\u8651\u5230\u9884\u8bad\u7ec3\u7684\u6210\u672c\uff0c\u8fd9\u4e9b\u7ed3\u679c\u8868\u660e\uff0c\u4e3a\u4e86\u5728\u591a\u6a21\u6001\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5b9e\u73b0\u76f8\u540c\u6027\u80fd\uff0c\u539f\u751f\u8bad\u7ec3\u53ef\u80fd\u662f\u66f4\u6709\u6548\u65b9\u6cd5\u3002<\/p>\n<p><strong>\u8fc8\u5411\u591a\u6a21\u6001\u4e13\u4e1a\u5316<\/strong><\/p>\n<p>\u7814\u7a76\u8bc1\u660e\u4e86\u5728\u56fa\u5b9a\u8ba1\u7b97\u9884\u7b97\u4e0b\uff0c\u65e9\u878d\u5408\u6a21\u578b\u7684\u6027\u80fd\u4e0e\u540e\u878d\u5408\u6a21\u578b\u76f8\u5f53\u3002\u7136\u800c\uff0c\u591a\u6a21\u6001\u6570\u636e\u672c\u8d28\u4e0a\u662f\u5f02\u6784\u7684\uff0c\u8bad\u7ec3\u4e00\u4e2a\u7edf\u4e00\u7684\u6a21\u578b\u6765\u62df\u5408\u5982\u6b64\u591a\u6837\u5316\u7684\u5206\u5e03\u53ef\u80fd\u5e76\u975e\u6700\u4f18\u65b9\u6848\u3002<\/p>\n<p>\u56e0\u6b64\u7814\u7a76\u4eba\u5458\u4e3b\u5f20\u5728\u7edf\u4e00\u67b6\u6784\u5185\u8fdb\u884c\u591a\u6a21\u6001\u7279\u5316\u3002\u7406\u60f3\u60c5\u51b5\u4e0b\uff0c\u6a21\u578b\u5e94\u8be5\u9690\u5f0f\u5730\u9002\u5e94\u6bcf\u79cd\u6a21\u6001\uff0c\u4f8b\u5982\uff0c\u901a\u8fc7\u5b66\u4e60\u7279\u5b9a\u4e8e\u6a21\u6001\u7684\u6743\u91cd\u6216\u4e13\u95e8\u7684\u4e13\u5bb6\u3002\u591a\u6a21\u6001\u6a21\u578b + MoE \u662f\u4e00\u4e2a\u6709\u6f5c\u529b\u7684\u65b9\u5411\uff0cMoE \u5df2\u5728 LLM \u4e2d\u8bc1\u660e\u4e86\u5176\u6709\u6548\u6027\u3002<\/p>\n<p>\u7814\u7a76\u89c2\u5bdf\u4e86\u4e0d\u540c\u6570\u91cf\u6d3b\u52a8\u53c2\u6570\u3001\u4e0d\u540c\u6570\u91cf token \u8bad\u7ec3\u7684\u6a21\u578b\uff0c\u6bd4\u8f83\u4e86\u7a00\u758f MoE \u6a21\u578b\u548c\u5bc6\u96c6 MoE \u6a21\u578b\u3002\u56fe 9 \u663e\u793a\uff0c\u5728\u76f8\u540c\u7684\u63a8\u7406\u6210\u672c\uff08\u6216\u6d3b\u52a8\u53c2\u6570\u6570\u91cf\uff09\u4e0b\uff0cMoE \u7684\u6027\u80fd\u663e\u8457\u4f18\u4e8e\u5bc6\u96c6\u6a21\u578b\u3002\u6709\u8da3\u7684\u662f\uff0c\u8fd9\u79cd\u6027\u80fd\u5dee\u8ddd\u5728\u6a21\u578b\u89c4\u6a21\u8f83\u5c0f\u65f6\u66f4\u4e3a\u660e\u663e\u3002\u8fd9\u8868\u660e MoE \u80fd\u591f\u4f7f\u6a21\u578b\u66f4\u6709\u6548\u5730\u5904\u7406\u5f02\u6784\u6570\u636e\uff0c\u5e76\u4e13\u6ce8\u4e8e\u4e0d\u540c\u7684\u6a21\u6001\u3002\u7136\u800c\uff0c\u968f\u7740\u5bc6\u96c6\u6a21\u578b\u89c4\u6a21\u53d8\u5f97\u8db3\u591f\u5927\uff0c\u4e24\u79cd\u67b6\u6784\u4e4b\u95f4\u7684\u5dee\u8ddd\u4f1a\u9010\u6e10\u7f29\u5c0f\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-084f3f166f6a7e51f656e4198bb7d6d9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-084f3f166f6a7e51f656e4198bb7d6d9.png\"><\/a><\/section>\n<p>\u5148\u524d\u5173\u4e8e\u5927\u6a21\u578b Scaling Laws \u7684\u7814\u7a76\u4e3b\u8981\u89c2\u5bdf\u9a8c\u8bc1\u635f\u5931\uff0c\u8fd9\u79cd\u8bc4\u4f30\u4e0e\u4e0b\u6e38\u4efb\u52a1\u7684\u8868\u73b0\u6709\u5f88\u597d\u7684\u76f8\u5173\u6027\u3002\u4e3a\u4e86\u9a8c\u8bc1\u8fd9\u4e00\u70b9\uff0c\u7814\u7a76\u4eba\u5458\u5bf9 LLaVA \u6df7\u5408\u6a21\u578b\u8fdb\u884c\u4e86\u591a\u6a21\u6001\u6307\u4ee4\u8c03\u6574\u9636\u6bb5 (SFT)\uff0c\u5e76\u62a5\u544a\u4e86\u5176\u5728\u591a\u4e2a VQA \u548c\u5b57\u5e55\u4efb\u52a1\u4e2d\u7684\u51c6\u786e\u7387\u548c CIDEr \u5f97\u5206\u3002\u8868 7 \u8bc1\u5b9e\u4e86\u4e0d\u540c\u6a21\u578b\u914d\u7f6e\u7684\u6392\u540d\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-89d7e0acf06aad2c565f9b64a194439e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-89d7e0acf06aad2c565f9b64a194439e.png\"><\/a><\/section>\n<p>\u5177\u4f53\u800c\u8a00\uff0c<strong>\u65e9\u878d\u5408\u4f18\u4e8e\u540e\u878d\u5408\uff0c\u800c\u591a\u6a21\u6001 MoE \u4f18\u4e8e\u5bc6\u96c6\u6a21\u578b<\/strong>\u3002\u7136\u800c\uff0c\u7531\u4e8e\u8fd9\u4e9b\u6a21\u578b\u89c4\u6a21\u76f8\u5bf9\u8f83\u5c0f\uff0815 \u4ebf\u89c4\u6a21\uff09\u3001\u4ece\u96f6\u5f00\u59cb\u8bad\u7ec3\uff0c\u5e76\u5728\u5c0f\u578b\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u5fae\u8c03\uff0c\u56e0\u6b64\u603b\u4f53\u5f97\u5206\u4f4e\u4e8e\u76ee\u524d\u7684 SOTA \u6c34\u5e73\u3002<\/p>\n<p>\u66f4\u591a\u7ec6\u8282\u53ef\u53c2\u770b\u8bba\u6587\u539f\u6587\u3002<\/p>\n<p><sup>\u53c2\u8003\u5185\u5bb9\uff1a<\/sup><\/p>\n<p><sup>https:\/\/x.com\/arankomatsuzaki\/status\/1910525957452292456<\/sup><\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-04-12-3\" target=\"_blank\">\u82f9\u679c\u53d1\u73b0\u539f\u751f\u591a\u6a21\u6001\u6a21\u578bScaling Laws\uff1a\u65e9\u878d\u5408\u4f18\u4e8e\u540e\u878d\u5408\uff0cMoE\u4f18\u4e8e\u5bc6\u96c6\u6a21\u578b<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u82f9\u679c\u53d1\u73b0\u539f\u751f\u591a\u6a21\u6001\u6a21\u578b [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[73,68],"class_list":["post-40098","post","type-post","status-publish","format-standard","hentry","category-news","tag-73","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u82f9\u679c\u53d1\u73b0\u539f\u751f\u591a\u6a21\u6001\u6a21\u578bScaling Laws\uff1a\u65e9\u878d\u5408\u4f18\u4e8e\u540e\u878d\u5408\uff0cMoE\u4f18\u4e8e\u5bc6\u96c6\u6a21\u578b - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=40098\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=40098\",\"url\":\"https:\/\/17aitech.com\/?p=40098\",\"name\":\"\u82f9\u679c\u53d1\u73b0\u539f\u751f\u591a\u6a21\u6001\u6a21\u578bScaling Laws\uff1a\u65e9\u878d\u5408\u4f18\u4e8e\u540e\u878d\u5408\uff0cMoE\u4f18\u4e8e\u5bc6\u96c6\u6a21\u578b - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=40098#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=40098#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dd8da943491889b92c64420bfe784029.png\",\"datePublished\":\"2025-04-12T17:02:45+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=40098#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=40098\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=40098#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dd8da943491889b92c64420bfe784029.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dd8da943491889b92c64420bfe784029.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=40098#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u82f9\u679c\u53d1\u73b0\u539f\u751f\u591a\u6a21\u6001\u6a21\u578bScaling Laws\uff1a\u65e9\u878d\u5408\u4f18\u4e8e\u540e\u878d\u5408\uff0cMoE\u4f18\u4e8e\u5bc6\u96c6\u6a21\u578b\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u82f9\u679c\u53d1\u73b0\u539f\u751f\u591a\u6a21\u6001\u6a21\u578bScaling Laws\uff1a\u65e9\u878d\u5408\u4f18\u4e8e\u540e\u878d\u5408\uff0cMoE\u4f18\u4e8e\u5bc6\u96c6\u6a21\u578b - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=40098","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=40098","url":"https:\/\/17aitech.com\/?p=40098","name":"\u82f9\u679c\u53d1\u73b0\u539f\u751f\u591a\u6a21\u6001\u6a21\u578bScaling Laws\uff1a\u65e9\u878d\u5408\u4f18\u4e8e\u540e\u878d\u5408\uff0cMoE\u4f18\u4e8e\u5bc6\u96c6\u6a21\u578b - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=40098#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=40098#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dd8da943491889b92c64420bfe784029.png","datePublished":"2025-04-12T17:02:45+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=40098#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=40098"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=40098#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dd8da943491889b92c64420bfe784029.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dd8da943491889b92c64420bfe784029.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=40098#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u82f9\u679c\u53d1\u73b0\u539f\u751f\u591a\u6a21\u6001\u6a21\u578bScaling Laws\uff1a\u65e9\u878d\u5408\u4f18\u4e8e\u540e\u878d\u5408\uff0cMoE\u4f18\u4e8e\u5bc6\u96c6\u6a21\u578b"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/40098","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=40098"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/40098\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=40098"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=40098"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=40098"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}