{"id":1226,"date":"2026-01-25T21:00:12","date_gmt":"2026-01-25T13:00:12","guid":{"rendered":"https:\/\/1ow.top\/index.php\/2026\/01\/25\/%e4%b8%8d%e8%af%bb%e5%8d%9a%e5%a3%ab%ef%bc%8c%e7%85%a7%e6%a0%b7%e8%bf%9bopenai%ef%bc%81o1%e6%a0%b8%e5%bf%83%e6%88%90%e5%91%98%e7%8e%b0%e8%ba%ab%e8%af%b4%e6%b3%95%e4%ba%86\/"},"modified":"2026-01-25T21:00:12","modified_gmt":"2026-01-25T13:00:12","slug":"%e4%b8%8d%e8%af%bb%e5%8d%9a%e5%a3%ab%ef%bc%8c%e7%85%a7%e6%a0%b7%e8%bf%9bopenai%ef%bc%81o1%e6%a0%b8%e5%bf%83%e6%88%90%e5%91%98%e7%8e%b0%e8%ba%ab%e8%af%b4%e6%b3%95%e4%ba%86","status":"publish","type":"post","link":"https:\/\/1ow.top\/index.php\/2026\/01\/25\/%e4%b8%8d%e8%af%bb%e5%8d%9a%e5%a3%ab%ef%bc%8c%e7%85%a7%e6%a0%b7%e8%bf%9bopenai%ef%bc%81o1%e6%a0%b8%e5%bf%83%e6%88%90%e5%91%98%e7%8e%b0%e8%ba%ab%e8%af%b4%e6%b3%95%e4%ba%86\/","title":{"rendered":"\u4e0d\u8bfb\u535a\u58eb\uff0c\u7167\u6837\u8fdbOpenAI\uff01o1\u6838\u5fc3\u6210\u5458\u73b0\u8eab\u8bf4\u6cd5\u4e86"},"content":{"rendered":"
\"\u4e0d\u8bfb\u535a\u58eb\uff0c\u7167\u6837\u8fdbOpenAI\uff01o1\u6838\u5fc3\u6210\u5458\u73b0\u8eab\u8bf4\u6cd5\u4e86\"\/<\/div><\/figure>\n
\"\u4e0d\u8bfb\u535a\u58eb\uff0c\u7167\u6837\u8fdbOpenAI\uff01o1\u6838\u5fc3\u6210\u5458\u73b0\u8eab\u8bf4\u6cd5\u4e86\"\/<\/div><\/figure>\n
\"\u4e0d\u8bfb\u535a\u58eb\uff0c\u7167\u6837\u8fdbOpenAI\uff01o1\u6838\u5fc3\u6210\u5458\u73b0\u8eab\u8bf4\u6cd5\u4e86\"\/<\/div><\/figure>\n
\"\u4e0d\u8bfb\u535a\u58eb\uff0c\u7167\u6837\u8fdbOpenAI\uff01o1\u6838\u5fc3\u6210\u5458\u73b0\u8eab\u8bf4\u6cd5\u4e86\"\/<\/div><\/figure>\n

henry \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n

\u5982\u679c\u6ca1\u6709PhD\uff0c\u662f\u4e0d\u662f\u5c31\u548c\u524d\u6cbfAI\u7814\u7a76\u6ca1\u5173\u7cfb\u4e86\uff1f<\/p>\n

\u81f3\u5c11\u5728Noam Brown\u770b\u6765\uff0c\u672a\u5fc5\u3002<\/p>\n

\u8fd9\u4f4dOpenAI \u7814\u7a76\u5458\u3001o1\u7684\u6838\u5fc3\u8d21\u732e\u8005\uff0c\u521a\u521a\u5206\u4eab\u4e86\u4e00\u4e32\u201c\u975e\u5178\u578b\u7814\u7a76\u5458\u201d\u7684\u7ecf\u5386\u3002<\/p>\n

\u6709\u4eba\u6ca1\u6709\u8bba\u6587\u3001\u6709\u4eba\u6ca1\u8bfb\u7814\u3001\u6709\u4eba\u767d\u5929\u5728\u9ea6\u80af\u9521\u4e0a\u73ed\uff0c\u665a\u4e0a\u5728GitHub\u5199\u7814\u7a76\u3002<\/p>\n

\u8fd8\u6709\u4eba\u6ca1\u4e8b\u5c31\u5728\u63a8\u7279\u53d1\u5e16\u3001GitHub\u4e0a\u63d0\u95ee\u7684\u3002<\/p>\n

\u4ed6\u4eec\u540e\u6765\u90fd\u53bb\u4e86\u54ea\uff1f<\/p>\n

OpenAI\u3001DeepMind\u3001Anthropic\u3002<\/p>\n

\u5728\u8fd9\u7bc7\u957f\u5e16\u4e2d\uff0cNoam\u5206\u4eab\u4e86Keller Jordan\u3001Sholto Douglas\u3001Andy Jones\u3001Kevin Wang\u7b49\u4e00\u6279\u201c\u975e\u5178\u578b\u201d\u7814\u7a76\u5458\u7684\u7ecf\u5386\u3002<\/p>\n

\u603b\u4f53\u770b\u4e0b\u6765\uff0c\u4ed6\u4eec\u90fd\u6709\u4e00\u4e9b\u76f8\u540c\u7684\u7279\u8d28\u3002\u6bd4\u5982\uff0c\u4e3b\u52a8\u6027\u6781\u5f3a\uff08\u70ed\u7231\uff09\u3001\u516c\u5f00\u7814\u7a76\uff0c\u4e0d\u95ed\u95e8\u9020\u8f66\u3001\u5de5\u7a0b\u80fd\u529b\u5728\u7ebf\u3001\u4f1aPR\u3001\u4e0d\u6267\u7740\u4e8e\u5934\u8854\u3002<\/p>\n

\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5177\u4f53\u6765\u770b\u3002<\/p>\n

\u9996\u5148\u88abNoam\u5206\u4eab\u7684\uff0c\u662fKeller Jordan\u3002<\/p>\n

\u4ed6\u73b0\u5728\u5728OpenAI\uff0c\u4ece\u4e8b\u9884\u8bad\u7ec3\u76f8\u5173\u7684\u7814\u7a76\u3002\u800c\u4ed6\u7684\u7814\u7a76\u751f\u6daf\uff0c\u67d0\u79cd\u7a0b\u5ea6\u4e0a\uff0c\u8d77\u4e8e\u4e00\u6bb5\u5178\u578b\u7684\u201c\u5957\u74f7\u7814\u7a76\u201d\u3002<\/p>\n

Keller\u53ea\u6709\u672c\u79d1\u5b66\u4f4d\uff08UCSD\uff09\uff0c\u4e5f\u6ca1\u6709\u4efb\u4f55\u8bba\u6587\u53d1\u8868\u7ecf\u5386\u3002<\/p>\n

\u6bd5\u4e1a\u540e\uff0c\u4ed6\u5148\u8fdb\u5165\u4e86\u4e00\u5bb6\u505a AI \u5185\u5bb9\u5ba1\u6838\u7684\u521d\u521b\u516c\u53f8\u5de5\u4f5c\u3002<\/p>\n

\u5728\u5de5\u4f5c\u4e4b\u4f59\uff0c\u4ed6\u5e76\u6ca1\u6709\u95f2\u7740\u3002<\/p>\n

Keller\u4e3b\u52a8\u8054\u7cfb\u4e86\u5f53\u65f6\u8fd8\u5728\u8c37\u6b4c\u7684Behnam Neyshabur\uff0c\u5e76\u5411\u4ed6\u5c55\u793a\u4e86\u4e00\u4e2a\u9488\u5bf9Behnam\u6700\u8fd1\u8bba\u6587\u7684\u6539\u8fdb\u60f3\u6cd5\u3002<\/p>\n

\u6216\u8bb8\u662f\u51fa\u4e8e\u201c\u6709\u4eba\u771f\u7684\u8ba4\u771f\u8bfb\u4e86\u6211\u8bba\u6587\u201d\u7684\u6210\u5c31\u611f\uff0cBehnam\u540c\u610f\u6307\u5bfc\u4ed6\u3002<\/p>\n

\u6700\u7ec8\uff0c\u4ed6\u4eec\u4e0e\u5176\u4ed6\u4f5c\u8005\u4e00\u8d77\uff0c\u5b8c\u6210\u4e86\u4e00\u7bc7\u5165\u9009ICLR 2023\u7684\u8bba\u6587\u3002<\/p>\n

\u5bf9\u4e8eKeller\u8fd9\u79cd\u201c\u5957\u74f7+\u7814\u7a76\u201d\u7684\u65b9\u5f0f\uff0cNoam\u8bc4\u4ef7\u8bf4\uff1a<\/p>\n

\u9057\u61be\u7684\u662f\uff0c\u5982\u4eca\u516c\u5f00\u7814\u7a76\u7684\u7a7a\u95f4\u6bd4\u4ee5\u524d\u5c0f\u4e86\uff0c\u4f46\u5728\u5df2\u6709\u8bba\u6587\u57fa\u7840\u4e0a\u63d0\u51fa\u6539\u8fdb\uff0c\u4f9d\u7136\u662f\u5411\u5b9e\u9a8c\u5ba4\u5185\u90e8\u7814\u7a76\u5458\u8bc1\u660e\u4f60\u80fd\u529b\u7684\u7edd\u4f73\u65b9\u5f0f\uff0c\u4e5f\u80fd\u8ba9\u5bf9\u65b9\u6709\u4fe1\u5fc3\u53bb\u5e2e\u4f60\u4e89\u53d6\u4e00\u6b21\u9762\u8bd5\u673a\u4f1a\u3002<\/p>\n

\u540e\u6765\uff0c\u4e00\u7bc72024\u5e74\u5e95\u5199\u4e0b\u7684\u4e00\u7bc7\u535a\u5ba2Muon: An optimizer for hidden layers in neural networks\uff0c\u5219\u771f\u6b63\u4e3aKeller\u62ff\u4e0b\u4e86\u4e00\u5f20\u53bbOpenAI\u7684\u8239\u7968\u3002<\/p>\n

\uff08\u6ce8\uff1aMuon\u662f\u4e13\u95e8\u7528\u4e8e\u795e\u7ecf\u7f51\u7edc\u9690\u85cf\u5c42\u7684\u4f18\u5316\u5668\uff0c\u6838\u5fc3\u601d\u60f3\u662f\u628a\u5e38\u89c4\u68af\u5ea6\u66f4\u65b0\u201c\u6b63\u4ea4\u5316\u201d\uff09<\/p>\n

\u5f53\u65f6\uff0cKeller\u53d1\u8d77\u5e76\u6301\u7eed\u66f4\u65b0\u4e86\u4e00\u4e2aNanoGPT speedrun\uff1a<\/p>\n

\u8fd9\u4e2a\u5b9e\u9a8c\u662f\u5728Karpathy\u7684NanoGPT\u6846\u67b6\u4e0b\uff0c\u7cfb\u7edf\u6027\u5730\u6bd4\u8f83\u4e0d\u540c\u8bad\u7ec3\u914d\u7f6e\u548c\u4f18\u5316\u5668\uff0c\u76ee\u6807\u662f\u628a\u9884\u8bad\u7ec3\u901f\u5ea6\u63a8\u5230\u6781\u9650\u3002<\/p>\n

\u800cKeller\u81ea\u5df1\u5012\u817e\u51fa\u7684Muon\uff0c\u6b63\u662f\u5728\u8fd9\u4e00\u7cfb\u5217speedrun\u5b9e\u9a8c\u4e2d\uff0c\u8131\u9896\u800c\u51fa\u3002<\/p>\n

\u8fd9\u5f15\u8d77\u4e86\u5927\u795eAndrej Karpathy\u7684\u5173\u6ce8\uff0cKeller\u4e5f\u56e0\u6b64\u8fdb\u5165\u4e86OpenAI\u7684\u89c6\u91ce\u3002<\/p>\n

\u800c\u540e\u7eed\u8fdb\u5165OpenAI\u4e5f\u662f\u987a\u624b\u63a8\u821f\u3002<\/p>\n

\u8fd9\u79cd\u628a\u7814\u7a76\u76f4\u63a5\u644a\u5728GitHub \u4e0a\u3001\u540c\u65f6\u5728\u793e\u4ea4\u5a92\u4f53\u4e0a\u6301\u7eed\u8bb0\u5f55\u7684\u65b9\u5f0f\uff0c\u4e00\u65b9\u9762\u8ba9\u4ed6\u7684\u7814\u7a76\u6c34\u5e73\u548c\u4ea7\u51fa\u53d8\u5f97\u5f02\u5e38\u5bb9\u6613\u8861\u91cf\uff1b<\/p>\n

\u53e6\u4e00\u65b9\u9762\uff0c\u4e5f\u8ba9\u8fd9\u4e9b\u5de5\u4f5c\u6ca1\u6709\u6df9\u6ca1\u5728arXiv\u7684\u8bba\u6587\u6d77\u91cc\uff0c\u800c\u662f\u88ab\u771f\u6b63\u770b\u89c1\uff0c\u5e76\u6700\u7ec8\u5f97\u5230\u4e86\u5708\u5185\u5927\u795e\u7684\u8ba4\u53ef\u3002<\/p>\n

\u8bf4\u5230\u5e95\uff0cPR\u592a\u91cd\u8981\u4e86\uff01<\/p>\n

\u63a5\u4e0b\u6765\uff0cNoam\u53c8\u5206\u4eab\u4e86\u73b0Anthropic\u7814\u7a76\u5458Sholto Douglas\u7684\u6545\u4e8b\u3002<\/p>\n

Sholto Douglas\u66fe\u7ecf\u5c31\u804c\u4e8e\u8c37\u6b4c\uff0c\u662fGemini\u9879\u76ee\u7684\u91cd\u8981\u63a8\u624b\u4e4b\u4e00\u3002<\/p>\n

\u800c\u4ed6\uff0c\u4e5f\u53ea\u6709\u672c\u79d1\u5b66\u4f4d\u3002<\/p>\n

\uff08\u6ca1\u9519\uff0c\u4ed6\u8fd8\u5728\u6e05\u534e\u4ea4\u6362\u4e86\u4e00\u5e74\uff09<\/p>\n

Sholto\u672c\u79d1\u5b66\u7684\u867d\u7136\u662f\u673a\u5668\u4eba\uff0c\u4f46\u6bd5\u4e1a\u540e\u5374\u8fdb\u4e86\u9ea6\u80af\u9521\u5de5\u4f5c\u3002<\/p>\n

\u800c\u8fd9\uff0c\u5f88\u5927\u7684\u4e00\u90e8\u5206\u539f\u56e0\u662f\uff0cSholto\u6ca1\u6709\u7533\u8bf7\u5230\u60f3\u53bb\u7684\u7814\u7a76\u751f\u9879\u76ee\u3002<\/p>\n

\u7531\u6b64\uff0cSholto\u7d22\u6027\u5e72\u8106\u4e00\u8fb9\u4e0a\u73ed\uff0c\u4e00\u8fb9\u505a\u7814\u7a76\u3002\u636eSholto\u5728\u64ad\u5ba2\u4e2d\u7684\u5206\u4eab\uff0c\u4ed6\u51e0\u4e4e\u662f\uff1a<\/p>\n

\u6bcf\u5929\u665a\u4e0a10\u70b9\u5230\u51cc\u66682\u70b9\uff0c\u5468\u672b\u6bcf\u5929\u81f3\u5c116\u52308\u4e2a\u5c0f\u65f6\uff0c\u5168\u90fd\u5728\u5199\u4ee3\u7801\u3001\u505a\u5b9e\u9a8c\u3001\u641e\u7814\u7a76\u9879\u76ee\u3002<\/p>\n

\u5728\u5b66\u4e60\u8fc7\u7a0b\u4e2d\uff0cSholto\u8fd8\u7ecf\u5e38\u5728Google\u5f00\u6e90\u7684\u673a\u5668\u5b66\u4e60\u6846\u67b6JAX\u7684\u5b98\u65b9GitHub\u4ed3\u5e93\u4e0a\u63d0\u95ee\u3002<\/p>\n

\u800c\u8fd9\uff0c\u5f15\u8d77\u4e86\u5f53\u65f6\u5728\u8c37\u6b4c\uff0c\u73b0\u5728\u5728Anthropic\u7684James Bradbury\u7684\u5174\u8da3\u3002<\/p>\n

James\u5fc3\u60f3\uff1a\u8fd9\u4e16\u754c\u4e0a\u8fd8\u6709\u6211\u4e0d\u8ba4\u8bc6\u7684\u9694\u8fd9\u63d0\u95ee\u7684\u4eba\uff1f<\/p>\n

\u4e8e\u662f\uff0c\u5728\u770b\u5230\u4e86Sholto\u7684\u63d0\u95ee\u4ee5\u53ca\u4ed6\u5e73\u65f6\u505a\u7684\u9879\u76ee\uff0cJames\u5c31\u4e3b\u52a8\u8054\u7cfb\u4ed6\uff0c\u5e76\u9080\u8bf7\u4ed6\u524d\u5f80Google DeepMind\u7684\u9762\u8bd5\u3002<\/p>\n

\u6700\u7ec8\uff0c\u672c\u79d1\u751f\uff0c\u7b2c\u4e00\u6bb5\u5de5\u4f5c\u5b8c\u5168\u4e0dAI\u7684Sholto\u8fdb\u5165\u4e86\u8c37\u6b4c\u3002<\/p>\n

\u6709\u8da3\u7684\u662f\uff0c\u5728\u8fdb\u5165\u8c37\u6b4c\u4e4b\u540e\uff0cSholto\u624d\u53d1\u73b0\uff0c\u81ea\u5df1\u88ab\u62db\u8fdb\u6765\uff0c\u5176\u5b9e\u5728\u67d0\u79cd\u7a0b\u5ea6\u4e0a\u662f\u4e00\u4e2a\u5185\u90e8\u5b9e\u9a8c\uff1a<\/p>\n

\u8bd5\u8bd5\u770b\uff0c\u80fd\u4e0d\u80fd\u628a\u4e00\u4e2a\u70ed\u60c5\u548c\u4e3b\u89c2\u80fd\u52a8\u6027\u6781\u5f3a\u7684\u4eba\uff0c\u548c\u4ed6\u4eec\u8ba4\u8bc6\u7684\u6700\u4f18\u79c0\u7684\u4e00\u6279\u5de5\u7a0b\u5e08\u914d\u5bf9\uff0c\u770b\u770b\u4f1a\u53d1\u751f\u4ec0\u4e48\u3002<\/p>\n

andy jones\u73b0\u5728\u662fAnthropic\u7684\u7814\u7a76\u5458\uff0c\u8d1f\u8d23\u7684RL\u57fa\u7840\u8bbe\u65bd\u3001\u6c99\u7bb1\u7cfb\u7edf\uff0c\u7275\u5934\u591a\u6a21\u6001\u65b9\u5411\uff0c\u4ee5\u53ca\u53c2\u4e0e\u5927\u6a21\u578b\u8bad\u7ec3\u3002<\/p>\n

\u5728\u6b64\u4e4b\u524d\uff0c\u4ed6\u662f\u4e00\u4f4d\u534a\u9000\u4f11\u7684\u91cf\u5316\u5206\u6790\u5e08\u3002<\/p>\n

\u548c\u4e0a\u9762\u51e0\u4f4d\u4e00\u6837\uff0candy\u6709\u6781\u5f3a\u7684\u4e3b\u52a8\u6027\uff1a\u4ed6\u81ea\u8d39\u79df\u8d41\u7b97\u529b\uff0c\u81ea\u8d39\u53d1\u8868\u8bba\u6587\u3002<\/p>\n

\u5728\u8bba\u6587\u300aScaling Scaling Laws with Board Games\u300b\u4e2d\uff0candy\u4e0d\u4ec5\u9996\u6b21\u7cfb\u7edf\u5730\u91cf\u5316train\u4e0etest compute\u7684tradeoff\u3002<\/p>\n

\u8fd8\u5728\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u6269\u5c55\u6210\u4e3a\u70ed\u95e8\u6982\u5ff5\u4e4b\u524d\uff0c\u6bd4\u8f83\u4e86\u6269\u5c55\u9884\u8bad\u7ec3\u548c\u6269\u5c55\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u7684\u5f71\u54cd\u3002<\/p>\n

\u8fd9\u7bc7\u8bba\u6587\u76f4\u63a5\u8ba9\u4ed6\u6210\u4e3a\u673a\u5668\u5b66\u4e60\u9886\u57df\u7684\u70b8\u5b50\u9e21\u3002<\/p>\n

\u800c\u540e\u6765\uff0c\u8fd9\u4e00\u7814\u7a76\u8fd8\u76f4\u63a5\u63a8\u52a8\u4e86o1\u7b49\u6a21\u578b\u7684\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u8303\u5f0f\uff0c\u5e76\u88abNoam\u653e\u5165\u4e86o1\u7684\u6280\u672f\u62a5\u544a\u91cc\u3002<\/p>\n

\u800c\u636e\u4e00\u4f4d\u7f51\u53cb\u7684\u8865\u5145\uff0candy\u7684\u7ecf\u5386\u5e76\u4e0d\u6b62\u5982\u6b64\u3002<\/p>\n

\u4ed6\u5728\u82f1\u56fd\u5b8c\u6210\u4e864\u5e74\u5236\u7684\u6570\u5b66\u7855\u58eb\uff0c\u5e76\u4e14\u82b19\u4e2a\u6708\u65f6\u95f4\u81ea\u5b66\u4e86\u8f6f\u4ef6\u5de5\u7a0b\u548c\u673a\u5668\u5b66\u4e60\u3002<\/p>\n

\u540e\u6765\uff0c\u4ed6\u53c8\u5e72\u4e86\u4e00\u5e74\u7684\u751f\u7269\u4fe1\u606f\u5b66\u7814\u7a76\u3002<\/p>\n

\u518d\u540e\u6765\uff0c3\u5e74\u7684\u91cf\u5316\u4ea4\u6613\u5de5\u4f5c\u5df2\u8db3\u4ee5\u8ba9\u4ed6\u8d22\u5bcc\u81ea\u7531\u3002<\/p>\n

\u5c3d\u7ba1\u5982\u6b64\uff0candy\u5e76\u6ca1\u6709\u505c\u4e0b\uff0c\u800c\u662f\u5b8c\u6210\u4e86\u4e00\u7cfb\u5217\u72ec\u7acb\u63a2\u7d22\uff0c\u5e76\u6700\u7ec8\u8fdb\u5165Anthropic\u3002<\/p>\n

\u6b64\u5916\uff0cNoam\u900f\u9732\uff0c\u50cfOpenAI\u8fd9\u6837\u7684\u5b9e\u9a8c\u5ba4\uff0c\u786e\u5b9e\u4f1a\u76f4\u63a5\u4ece\u672c\u79d1\u9636\u6bb5\u62db\u8058\u7814\u7a76\u5458\u3002<\/p>\n

\u800cKevin Wang\u5c31\u662f\u5176\u4e2d\u4e00\u3002<\/p>\n

2025\u5e74\uff0cKevin Wang\u5728\u5b8c\u6210\u672c\u79d1\u5b66\u4e60\u540e\uff0c\u76f4\u63a5\u8fdb\u5165OpenAI\uff0c\u4ece\u672c\u79d1\u751f\u4e00\u6b65\u8de8\u5165\u4e00\u7ebf\u7814\u7a76\u56e2\u961f\u3002<\/p>\n

\u8fd9\u5e76\u4e0d\u662f\u4e00\u6b21\u201c\u7834\u683c\u7684\u8fd0\u6c14\u201d\u3002<\/p>\n

\u4e00\u65b9\u9762\uff0cWang\u62e5\u6709\u5bfc\u5e08\u7684\u5f3a\u529b\u63a8\u8350\uff1b\u4f46\u66f4\u5173\u952e\u7684\u662f\uff0c\u4ed6\u672c\u8eab\u5c31\u63e1\u7740\u4e00\u7bc7\u6781\u5176\u624e\u773c\u7684\u8bba\u6587\u3002<\/p>\n

\u5bf9\u4e8eOpenAI\u6765\u8bf4\uff0c\u9876\u4f1a\u4e0a\u201c\u6c34\u5e73\u8fd8\u4e0d\u9519\u201d\u7684\u8bba\u6587\u5e76\u4e0d\u7a00\u7f3a\u3002\u771f\u6b63\u7a00\u7f3a\u7684\uff0c\u662f\u90a3\u79cd\u5728\u7b2c\u4e00\u773c\u5c31\u80fd\u88ab\u6311\u51fa\u6765\u7684\u5de5\u4f5c\u3002<\/p>\n

Noam\u56de\u5fc6\uff0c\u5728NeurIPS 2025\u76845290\u7bc7\u8bba\u6587\u4e2d\uff0c\u4ed6\u4eec\u51e0\u4e4e\u662f\u4e00\u773c\u5c31\u6ce8\u610f\u5230\u4e86Wang\u7684\u7814\u7a76\u3002<\/p>\n

\u800c\u540e\u6765\u7684\u7ed3\u679c\uff0c\u4e5f\u9a8c\u8bc1\u4e86\u8fd9\u4e2a\u5224\u65ad\u2014\u2014<\/p>\n

\u4ed6\u7684\u7b2c\u4e00\u4f5c\u8005\u8bba\u6587\u300a1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities\u300b\uff0c\u6700\u7ec8\u62ff\u4e0b\u4e86NeurIPS 2025\u6700\u4f73\u8bba\u6587\u3002<\/p>\n

\u8c08\u5230\u62db\u8058\u6807\u51c6\u65f6\uff0cNoam \u4e5f\u7ed9\u51fa\u4e86\u4e00\u4e2a\u5f88\u73b0\u5b9e\u7684\u8bc4\u4ef7\uff1a<\/p>\n

\u5bfc\u5e08\u7684\u63a8\u8350\u5206\u91cf\u5f88\u91cd\u3002\u56e0\u4e3a\u4ec5\u51ed\u7b80\u5386\uff0c\u751a\u81f3\u53ea\u770b\u8bba\u6587\uff0c\u672c\u8eab\u5c31\u5f88\u96be\u5168\u9762\u8bc4\u4f30\u4e00\u540d\u7814\u7a76\u5458\u7684\u771f\u5b9e\u6c34\u5e73\u3002<\/p>\n

\u6700\u540e\uff0cNoam\u4e5f\u5206\u4eab\u4e86\u81ea\u5df1\u5bf9\u201c\u85aa\u916c\u201d\u7684\u770b\u6cd5\uff1a<\/p>\n

\u5173\u4e8e\u201c\u85aa\u916c\u201d\u8fd9\u4ef6\u4e8b\uff1a\u6211\u8ba4\u8bc6\u4e00\u4e9b\u4eba\u505a\u91cf\u5316\u662f\u4e3a\u4e86\u8d5a\u94b1\uff0c\u4f46\u4e94\u5e74\u4e4b\u540e\u5374\u5f00\u59cb\u53cd\u95ee\u81ea\u5df1\u2014\u2014\u6211\u5230\u5e95\u5728\u7528\u4eba\u751f\u5e72\u4ec0\u4e48\uff1f<\/p>\n

\u6211\u4eec\u6b63\u5904\u5728\u4e00\u4e2a\u975e\u5e38\u7279\u6b8a\u7684\u5386\u53f2\u65f6\u523b\u3002\u505aAI\u7814\u7a76\uff0c\u4f60\u65e2\u6709\u673a\u4f1a\u79ef\u6781\u53c2\u4e0e\u5e76\u5f15\u5bfc\u8fd9\u4e2a\u65f6\u4ee3\u6700\u91cd\u8981\u7684\u6280\u672f\u53d1\u5c55\uff0c\u540c\u65f6\u6536\u5165\u4e5f\u76f8\u5f53\u53ef\u89c2\u3002<\/p>\n

\u6211\u672c\u79d1\u6bd5\u4e1a\u540e\u5728\u91cf\u5316\u4ea4\u6613\u884c\u4e1a\u5e72\u4e86\u4e00\u5e74\uff0c\u4f46\u5e76\u4e0d\u5e0c\u671b\u81ea\u5df1\u6b64\u751f\u5bf9\u4eba\u7c7b\u7684\u201c\u8d21\u732e\u201d\uff0c\u53ea\u662f\u8ba9\u80a1\u7968\u5e02\u573a\u7a0d\u5fae\u66f4\u6709\u6548\u7387\u4e00\u70b9\u3002<\/p>\n

\u540e\u6765\u6211\u9009\u62e9\u964d\u85aa\u8f6c\u53bb\u505a AI \u7814\u7a76\u2014\u2014\u8fd9\u662f\u6211\u8fd9\u8f88\u5b50\u505a\u8fc7\u6700\u597d\u7684\u51b3\u5b9a\u3002\u653e\u5728\u4eca\u5929\uff0c\u4f60\u751a\u81f3\u4e0d\u9700\u8981\u964d\u85aa\uff0c\u4e5f\u80fd\u8d70\u4e0a\u8fd9\u6761\u8def\u3002<\/p>\n

\u9664\u4e86Noam\u7684\u5206\u4eab\u4ee5\u5916\uff0c\u73b0\u5728\u53ef\u80fd\u8d8a\u6765\u8d8a\u660e\u663e\u7684\u4e00\u70b9\u662f\uff1a<\/p>\n

\u60f3\u628a\u7814\u7a76\u505a\u597d\u3001\u60f3\u8fdb\u524d\u6cbfAI\u5927\u5382\uff0c\u535a\u58eb\u5b66\u4f4d\u672c\u8eab\uff0c\u4f3c\u4e4e\u5df2\u7ecf\u4e0d\u518d\u662f\u5fc5\u9700\u54c1\u4e86\u3002<\/p>\n

\u6bd4\u5982\uff0cGPT\u8bba\u6587\u7684\u4e00\u4f5c\u3001CLIP\u7684\u4e00\u4f5cAlec Radford\uff0c\u5c31\u662f\u672c\u79d1\u51fa\u8eab\u3002<\/p>\n

\u800c\u73b0\u5728OpenAI\u7684\u9996\u5e2d\u7814\u7a76\u5458Mark Chen\uff0c\u4e5f\u53ea\u6709\u672c\u79d1\u5b66\u5386\uff0c\u4e00\u5f00\u59cb\u505a\u7684\u751a\u81f3\u8fd8\u662f\u91cf\u5316\u4ea4\u6613\u3002<\/p>\n

\u5728 Noam \u90a3\u6761\u5e16\u5b50\u4e0b\u9762\uff0c\u6709\u7f51\u53cb\u4e00\u9488\u89c1\u8840\u5730\u603b\u7ed3\u8fc7\uff1a<\/p>\n

\u4e00\u6761\u5f88\u68d2\u7684\u957f\u5e16\uff0c\u8bb2\u7684\u662f\u6ca1\u6709PhD\u5b66\u4f4d\u7684\u4eba\uff0c\u662f\u5982\u4f55\u8fdb\u5165\u524d\u6cbfAI\u5b9e\u9a8c\u5ba4\u505a\u7814\u7a76\u7684\u3002PhD\u80fd\u7ed9\u4f60\u5bfc\u5e08\u548c\u540c\u884c\u5708\u5b50\uff0c\u4f46\u5b83\u5e76\u4e0d\u4f1a\u81ea\u52a8\u8d4b\u4e88\u4f60\u597d\u5947\u5fc3\u3001\u4e3b\u52a8\u6027\uff0c\u6216\u8005\u7814\u7a76\u54c1\u5473\u3002<\/p>\n

\u4e0a\u9762\u63d0\u5230\u7684Keller\u5c31\u662f\u4e00\u4e2a\u5178\u578b\u4f8b\u5b50\u3002<\/p>\n

\u4ed6\u662f\u5728\u5b8c\u5168\u516c\u5f00\u7684\u73af\u5883\u91cc\u505a\u51fa\u4e86Muon\u4f18\u5316\u5668\uff0c\u5b9e\u9a8c\u8fc7\u7a0b\u3001\u7ed3\u679c\u90fd\u76f4\u63a5\u53d1\u5728\u63a8\u7279\u4e0a\u3002<\/p>\n

\u751a\u81f3\u6ca1\u6709\u8bba\u6587\uff0c\u53ea\u6709\u4e00\u7bc7\u535a\u5ba2\u3002<\/p>\n

\u4f46\u5f71\u54cd\u529b\u4e0d\u8a00\u81ea\u660e\u2014\u2014\u73b0\u5728OpenAI\u3001Kimi\u3001DeepSeek\u90fd\u5728\u4f7f\u7528\u5b83\u3002<\/p>\n

\u7c7b\u4f3c\u7684\u6545\u4e8b\u5e76\u4e0d\u5c11\u89c1\u3002<\/p>\n

\u6bd4\u5982Stability AI\u768480\u540d\u7814\u7a76\u5458\u548c\u5de5\u7a0b\u5e08\u4e2d\uff0c\u53ea\u670916\u4f4d\u62e5\u6709PhD\uff0c\u800c\u4e14\u5176\u4e2d\u76f8\u5f53\u4e00\u90e8\u5206\u662f\u76f4\u63a5\u4ece\u63a8\u7279\u4e0a\u62db\u6765\u7684\u3002<\/p>\n

\u6362\u53e5\u8bdd\u8bf4\uff0c\u4f60\u4e0d\u9700\u8981 PhD\uff0c\u624d\u80fd\u6210\u4e3a\u4e00\u540d\u4f18\u79c0\u7684\u7814\u7a76\u5458\u6216\u5de5\u7a0b\u5e08\u3002\u4f60\u53ea\u9700\u8981\u771f\u7684\u53bb\u505a\u4e8b\uff0c\u800c\u4e14\u8ba9\u522b\u4eba\u770b\u5f97\u89c1\u4f60\u5728\u505a\u4ec0\u4e48\u3002<\/p>\n

\u524d\u51e0\u5929\u521a\u4e0a\u300aWhyNotTV\u300b\u7684\u5609\u5bbe\u2014\u2014<\/p>\n

\u4eceGPT-3.5\u5230GPT-5\u90fd\u6709\u6df1\u5ea6\u53c2\u4e0e\u3001\u8d1f\u8d23GPT post-training RL infrastructure\u7684\u6838\u5fc3\u6210\u5458\u7fc1\u5bb6\u7fcc\uff0c\u4e5f\u662f\u53ea\u6709\u7855\u58eb\u5b66\u4f4d\u3002<\/p>\n

\u5728\u64ad\u5ba2\u91cc\u4ed6\u4e5f\u63d0\u5230\uff1a\u76f8\u6bd4\u53bb\u8bfb\u4e00\u4e2a PhD\uff0c\u5c3d\u65e9\u8fdb\u5165\u5de5\u4e1a\u754c\u6bd4\u8f83\u597d\u3002<\/p>\n

\u56e0\u4e3a\u7b49PhD\u8bfb\u5b8c\uff0c\u7814\u7a76\u8303\u5f0f\u53ef\u80fd\u5df2\u7ecf\u6362\u4e86\u4e00\u8f6e\uff1b<\/p>\n

\u800c\u4e14\u535a\u58eb\u9636\u6bb5\u5f80\u5f80\u9762\u5bf9\u7684\u662ftoy benchmark\uff0c\u800c\u4e0d\u662f\u771f\u5b9e\u4e16\u754c\u7684\u7cfb\u7edf\u548c\u7ea6\u675f\u3002<\/p>\n

\u4e0e\u6b64\u540c\u65f6\uff0c\u5f88\u591aAI Lab\u771f\u6b63\u77ed\u7f3a\u7684\uff0c\u53cd\u800c\u662f\u5de5\u7a0b\u80fd\u529b\u6781\u5f3a\u7684\u4eba\u3002<\/p>\n

\u81f3\u4e8e\u80fd\u4e0d\u80fd\u8fdb\u3001\u8fdb\u4e86\u4e4b\u540e\u505a\u4ec0\u4e48\uff0c\u5f88\u5927\u7a0b\u5ea6\u4e0a\u53d6\u51b3\u4e8e\uff1a\u5b9e\u9a8c\u5ba4\u5f53\u4e0b\u9700\u8981\u4ec0\u4e48\uff0c\u4ee5\u53ca\u4f60\u662f\u5426\u5df2\u7ecf\u5728\u8fc7\u5f80\u7ecf\u5386\u4e2d\u8bc1\u660e\u8fc7\u81ea\u5df1\u80fd\u8865\u4e0a\u90a3\u5757\u62fc\u56fe\u3002<\/p>\n

\u53c2\u8003\u94fe\u63a5[1]https:\/\/x.com\/polynoamial\/status\/2014084431062114744[2]https:\/\/www.youtube.com\/channel\/UC5xLV_gJAP9psKcyrJ3ZIcw[3]https:\/\/www.marketingaiinstitute.com\/blog\/the-marketing-ai-show-episode-74[4]https:\/\/www.dwarkesh.com\/p\/sholto-douglas-trenton-bricken<\/p>\n

\n

\u7248\u6743\u58f0\u660e\uff1a<\/strong><\/p>\n

\u672c\u535a\u5ba2\u90e8\u5206\u5185\u5bb9\u4e3a\u8f6c\u8f7d\u6587\u7ae0\uff0c\u65e8\u5728\u5206\u4eab\u6709\u4ef7\u503c\u7684\u4fe1\u606f\uff0c\u7248\u6743\u5f52\u539f\u4f5c\u8005\u6240\u6709\u3002<\/p>\n

\u8f6c\u8f7d\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u4e0e\u4ea4\u6d41\u76ee\u7684\uff0c\u4e0d\u5bf9\u6587\u7ae0\u89c2\u70b9\u8d1f\u8d23\uff0c\u4ea6\u4e0d\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u7528\u9014\u3002<\/p>\n

\u5982\u6d89\u53ca\u7248\u6743\u95ee\u9898\uff0c\u8bf7\u8054\u7cfb\u672c\u4eba\u5220\u9664\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"

henry \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n

\u5982\u679c\u6ca1\u6709PhD\uff0c\u662f\u4e0d\u662f\u5c31\u548c\u524d\u6cbfAI\u7814\u7a76\u6ca1\u5173\u7cfb\u4e86\uff1f<\/p>\n

\u81f3\u5c11\u5728Noam Brown\u770b\u6765\uff0c\u672a\u5fc5\u3002<\/p>\n

\u8fd9\u4f4dOpenAI \u7814\u7a76\u5458\u3001o1\u7684\u6838\u5fc3\u8d21\u732e\u8005\uff0c\u521a\u521a\u5206\u4eab\u4e86\u4e00\u4e32\u201c\u975e\u5178\u578b\u7814\u7a76\u5458\u201d\u7684\u7ecf\u5386\u3002<\/p>\n

\u6709\u4eba\u6ca1\u6709\u8bba\u6587\u3001\u6709\u4eba\u6ca1\u8bfb\u7814\u3001\u6709\u4eba\u767d\u5929\u5728\u9ea6\u80af\u9521\u4e0a\u73ed\uff0c\u665a\u4e0a\u5728GitHub\u5199\u7814\u7a76\u3002<\/p>\n

\u8fd8\u6709\u4eba\u6ca1\u4e8b\u5c31\u5728\u63a8\u7279\u53d1\u5e16\u3001GitHub\u4e0a\u63d0\u95ee\u7684\u3002<\/p>\n

\u4ed6\u4eec\u540e\u6765\u90fd\u53bb\u4e86\u54ea\uff1f<\/p>\n

OpenAI\u3001DeepMind\u3001Anthr<\/p>\n","protected":false},"author":1,"featured_media":1222,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[103],"class_list":["post-1226","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-xwzx","tag-openai"],"_links":{"self":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/1226","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/comments?post=1226"}],"version-history":[{"count":0,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/1226\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media\/1222"}],"wp:attachment":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media?parent=1226"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/categories?post=1226"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/tags?post=1226"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}