{"id":48027,"date":"2024-01-26T17:48:26","date_gmt":"2024-01-26T09:48:26","guid":{"rendered":"https:\/\/swarma.org\/?p=48027"},"modified":"2024-01-26T17:48:26","modified_gmt":"2024-01-26T09:48:26","slug":"yoshua-bengio%e6%bc%94%e8%ae%b2%e3%80%81neurips%e5%af%b9%e9%bd%90%e5%b7%a5%e4%bd%9c%e5%9d%8a%ef%bc%9a%e9%81%bf%e5%85%8dai%e7%81%be%e9%9a%be%e6%80%a7%e9%a3%8e%e9%99%a9%e7%9a%84%e6%8a%80%e6%9c%af","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=48027","title":{"rendered":"Yoshua Bengio\u6f14\u8bb2\u3001NeurIPS\u5bf9\u9f50\u5de5\u4f5c\u574a\uff1a\u907f\u514dAI\u707e\u96be\u6027\u98ce\u9669\u7684\u6280\u672f\u4e0e\u534f\u8c03\u6311\u6218"},"content":{"rendered":"
\n
\n
\n
\n
\n
\n
\n
\n

<\/p>\n<\/section>\n

\n
\n
\n
\n
\n
\n

<\/section>\n<\/section>\n
\n
\n

\u5bfc\u8bed<\/strong><\/p>\n<\/section>\n<\/section>\n

\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\n
\n

\u5728 2023 \u5e74 12 \u6708\uff0c\u7b2c\u4e8c\u5c4a AI \u5bf9\u9f50\u5de5\u4f5c\u574a\uff08Alignment Workshop\uff09\u5728\u65b0\u5965\u5c14\u826f\u7684 NeurIPS \u4f1a\u8bae\u671f\u95f4\u4e3e\u529e\u3002\u672c\u6b21\u5de5\u4f5c\u574a\u662f\u7531 Adam Gleave \u9886\u5bfc\u7684\u975e\u76c8\u5229\u6027\u7814\u7a76\u673a\u6784 FAR AI \u4e3b\u529e\u3002<\/span><\/strong><\/p>\n


<\/span><\/strong><\/p>\n

\u5de5\u4f5c\u574a\u9080\u8bf7\u5230\u4e86\u6765\u81ea\u5de5\u4e1a\u754c\u548c\u5b66\u672f\u754c150\u4f59\u4f4dAI\u7814\u7a76\u8005\u5c31 AI \u5b89\u5168\u548c\u5bf9\u9f50\u76f8\u5173\u7684\u7814\u7a76\u4e3b\u9898\u5c55\u5f00\u8ba8\u8bba\u548c\u8fa9\u8bba\uff0c\u4ece\u800c\u66f4\u597d\u5730\u7406\u89e3\u524d\u6cbf AI \u53ef\u80fd\u5e26\u6765\u7684\u98ce\u9669\uff0c\u5e76\u5bfb\u627e\u964d\u4f4e\u8fd9\u4e9b\u98ce\u9669\u7684\u7b56\u7565\u3002\u5de5\u4f5c\u574a\u7684\u8bb2\u8005\u548c\u53c2\u4e0e\u8005\u6709\u6765\u81eaOpenAI\u3001Anthropic\u3001Google DeepMind\u7b49\u9876\u5c16\u4e1a\u754cAI\u5b9e\u9a8c\u5ba4\u7684AGI\u5b89\u5168\u56e2\u961f\u6210\u5458\uff0c\u4e5f\u6709\u6765\u81eaMIT\u3001UC Berkeley\u3001CMU\u3001\u5251\u6865\u5927\u5b66\u3001\u725b\u6d25\u5927\u5b66\u3001Mila\u7b49\u9876\u5c16\u9ad8\u6821\u7684\u5b66\u8005\u3002\u56fe\u7075\u5956\u5f97\u4e3bYoshua Bengio\u5728\u5de5\u4f5c\u574a\u4e0a\u505a\u4e86\u4e3b\u65e8\u6f14\u8bb2\u3002<\/span><\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

\n
\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n

\u5b89\u8fdcAI<\/span> | \u6765\u6e90<\/strong><\/span><\/strong><\/p>\n


<\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n


<\/section>\n
<\/p>\n

\u65b0\u5965\u5c14\u826f\u5bf9\u9f50\u5de5\u4f5c\u574a\uff08New Orleans Alignment Workshop\uff09<\/span><\/p>\n


<\/span><\/section>\n<\/figure>\n
\u6574\u4e2a\u5de5\u4f5c\u574a\u6309\u7167AGI\u5b89\u5168\u7684\u7814\u7a76\u65b9\u5411\u5206\u6210\u4e86\u4e94\u4e2a\u4e3b\u9898\uff1a\u5bf9\u9f50\u95ee\u9898\u5bfc\u8bba<\/span>\uff08Introducing Alignment Problems\uff09<\/span>\u3001\u76d1\u7763<\/span>\uff08Oversight\uff09<\/span>\u3001\u9c81\u68d2\u6027\u548c\u6cdb\u5316<\/span>\uff08Robustness and Generalization\uff09<\/span>\u3001\u53ef\u89e3\u91ca\u6027<\/span>\uff08Interpretability\uff09<\/span>\u3001\u6cbb\u7406<\/span>\uff08Governance\uff09<\/span>\uff0c\u4ee5\u4e0b\u662f\u6240\u6709\u8bb2\u8005\u4e0e\u6f14\u8bb2\u4e3b\u9898\u4e00\u89c8<\/span>\uff08\u52a0\u7c97\u4e3a\u957f\u6f14\u8bb2\uff0c\u5176\u4f59\u4e3a\u77ed\u6f14\u8bb2\uff09<\/span>\uff1a<\/span><\/section>\n

<\/span><\/section>\n
<\/figcaption><\/figure>\n
<\/figcaption><\/figure>\n
<\/figure>\n

<\/section>\n
\u5b89\u8fdc AI \u7684\u6280\u672f\u9879\u76ee\u7ecf\u7406\u6bb5\u96c5\u6587\u53d7\u9080\u53c2\u4e0e\u4e86\u672c\u6b21\u5de5\u4f5c\u574a\u3002\u672c\u6587\u662f\u5de5\u4f5c\u574a\u603b\u7ed3\u7cfb\u5217\u63a8\u6587\u7684\u7b2c\u4e00\u7bc7\u6587\u7ae0\u3002\u603b\u7ed3\u4e86\u6765\u81eaYoshua Bengio<\/span>\uff08Mila\uff09<\/span>\u3001Adam Gleave<\/span>\uff08FAR AI\uff09<\/span>\u548cOwain Evans<\/span>\uff08Oxford\uff09<\/span>\u7684\u4e3b\u65e8\u6f14\u8bb2\u3002<\/span><\/section>\n

<\/section>\n


<\/p>\n

\n
<\/section>\n<\/section>\n

PART 1 \u4e3b\u65e8\u6f14\u8bb2 – Yoshua Bengio: Towards Quantitative Safety Guarantees and Alignment<\/span><\/strong><\/h1>\n

<\/section>\n
<\/p>\n
Yoshua Bengio: Towards Quantitative Safety Guarantees and Alignment<\/span><\/section>\n

<\/span><\/section>\n<\/figure>\n
\u5173\u4e8e\u4ec0\u4e48\uff1a\u6df1\u5ea6\u5b66\u4e60\u4e09\u5de8\u5934\u4e4b\u4e00\u3001\u56fe\u7075\u5956\u5f97\u4e3b Yoshua Bengio \u5728\u5de5\u4f5c\u574a\u53d1\u8868\u4e86\u4e3b\u65e8\u6f14\u8bb2\uff0c\u8bb2\u8ff0\u4e86\u4ed6\u5bf9AI\u98ce\u9669\u7684\u6700\u65b0\u7814\u5224\u548c\u4ed6\u6b63\u5728\u5b9e\u8df5\u7684AI\u5b89\u5168\u6280\u672f\u65b9\u5411 – \u201c\u968f\u7740\u7b97\u529b\u589e\u52a0\u800c\u589e\u5f3a\u7684\u5b89\u5168\u4fdd\u8bc1\u201d\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u98ce\u9669\u7814\u5224\uff1aBengio\u8ba4\u4e3a\u5f53\u524d\u907f\u514dAI\u707e\u96be\u6027\u98ce\u9669\u81f3\u5c11\u5b58\u5728\u7684\u4e24\u4e2a\u6311\u6218\uff1a\uff081\uff09\u89e3\u51b3\u5bf9\u9f50\u4e0e\u63a7\u5236\u6311\u6218\uff1a\u4e00\u4e2a\u6280\u672f\u6311\u6218+\u653f\u6cbb\u6311\u6218\uff08\u9700\u8981\u5927\u91cf\u7684\u5b89\u5168\u9886\u57df\u7684\u7814\u53d1\u6295\u8d44\uff09\uff1b\uff082\uff09\u89e3\u51b3\u534f\u8c03\u6311\u6218\uff1a\u786e\u4fddAI\u7814\u53d1\u8005\u9075\u5faa\u5b89\u5168\u548c\u4f26\u7406\u534f\u8bae\uff0c\u5e76\u4e14\u505a\u51fa\u5e94\u5bf9\u5931\u63a7AI\u7684\u51c6\u5907\u63aa\u65bd\uff0c\u540c\u65f6\u4e5f\u662f\u4e00\u4e2a\u793e\u4f1a\u6280\u672f\u6311\u6218\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/p>\n
\u907f\u514dAI\u707e\u96be\u6027\u98ce\u9669\u7684\u4e24\u4e2a\u6311\u6218<\/span><\/section>\n

<\/span><\/section>\n<\/figure>\n
Bengio\u7684\u6f14\u8bb2\u524d\u534a\u90e8\u5206\u4e3b\u8981\u8bb2\u8ff0\u4e86\u5982\u4f55\u4ece\u6cbb\u7406\u7684\u89c6\u89d2\u89e3\u51b3\u534f\u8c03\u6311\u6218\uff0c\u800c\u540e\u534a\u90e8\u5206\u8bb2\u8ff0\u4e86\u9488\u5bf9\u6280\u672f\u6311\u6218\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u672c\u6587\u5c06\u96c6\u4e2d\u8ba8\u8bba\u6280\u672f\u6311\u6218\u7684\u90e8\u5206\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u98ce\u9669\u6a21\u578b\uff1aBengio\u5728\u8bb2\u5ea7\u4e2d\u5f3a\u8c03\u7684\u98ce\u9669\u6a21\u578b\u4e3b\u8981\u662f\u5956\u52b1\u7834\u89e3<\/span>\uff08reward hacking\uff09<\/span>\u3002\u5047\u8bbe\u6a21\u578b\u4f30\u8ba1\u4e86\u4e00\u4e2a\u4e0d\u5b8c\u7f8e\u7684\u5956\u52b1\u51fd\u6570\uff0c\u5e76\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u7684\u65b9\u6cd5\u5bf9\u5176\u8fdb\u884c\u4f18\u5316\uff0c\u6a21\u578b\u4f1a\u66f4\u6709\u53ef\u80fd\u627e\u5230\u66f4\u5927\u7684\u771f\u5b9e\u5956\u52b1\u51fd\u6570\u548c\u4f30\u8ba1\u5956\u52b1\u51fd\u6570\u7684\u4e0d\u540c<\/span>\uff08mismatch\uff09<\/span>\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/p>\n
\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u5bfc\u81f4\u7684\u5bf9\u9f50\u5931\u8d25<\/span><\/section>\n

<\/span><\/section>\n<\/figure>\n
\u5f53\u524d\u7684\u8303\u5f0f\u7684\u7f3a\u9677\u5728\u4e8e\uff1a\u6a21\u578b\u80fd\u529b\u8d8a\u5f3a\uff0c\u6a21\u578b\u8fdb\u884c\u5956\u52b1\u7834\u89e3\u9020\u6210\u707e\u96be\u7684\u80fd\u529b\u4f1a\u8d8a\u5f3a\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff1a\u5982\u679cAI\u80fd\u591f\u201c\u7406\u89e3\u201d\u5956\u52b1\u6765\u6e90\u4e8e\u4eba\u7c7b\u5728\u952e\u76d8\u4e0a\u8f93\u5165+1\u7684\u52a8\u4f5c\uff0c\u90a3\u4e48AI\u5c31\u6709\u7406\u7531\u63a5\u7ba1\u4eba\u7c7b\u7684\u952e\u76d8\u800c\u8f93\u5165\u5f88\u591a+1\u3002\u53e6\u4e00\u4e2a\u65b9\u9762\uff0c\u5047\u8bbeAI\u5b66\u5230\u4e86\u8fd9\u6837\u7684\u7b56\u7565\uff0c\u4e3a\u4e86\u4fdd\u8bc1\u5176\u5956\u52b1\u7684\u6700\u5927\u5316\uff0c\u5b83\u4e5f\u6709\u7406\u7531\u6b3a\u9a97\u4eba\u7c7b\u800c\u4e0d\u5e0c\u671b\u4eba\u7c7b\u53d1\u73b0\u5176\u5b66\u5230\u7684\u7b56\u7565\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/p>\n
RL+\u7b97\u529b\u5982\u4f55\u5bfc\u81f4\u5956\u52b1\u7834\u89e3<\/span><\/section>\n

<\/span><\/section>\n<\/figure>\n
\u4e00\u79cd\u5e94\u5bf9\u7684\u65b9\u6cd5\u662f\u300aHuman Compatible\u300b\u7684\u4f5c\u8005Stuart Russell\u63d0\u51fa\u7684\u5408\u4f5c\u9006\u5f3a\u5316\u5b66\u4e60<\/span>\uff08Cooperative Inverse RL\uff1bCIRL\uff09<\/span>\uff0c\u5373\u4e0d\u53ea\u662f\u5b66\u5230\u5355\u4e00\u5956\u52b1\u51fd\u6570\uff0c\u800c\u662f\u7ed9\u5b9a\u5df2\u6709\u7684\u4fe1\u606f\u5b66\u4e60\u5956\u52b1\u51fd\u6570\u7684\u5206\u5e03\u6765\u4fdd\u6301\u5bf9\u4eba\u7c7b\u504f\u597d\u7684\u4e0d\u786e\u5b9a\u6027\u3002\u8fd9\u6837\u7684\u65b9\u6cd5\u542f\u53d1\u4e86Bengio\u6b63\u5728\u5b9e\u8df5\u7684\u53e6\u4e00\u79cd\u8303\u5f0f\uff1a\u5373\u901a\u8fc7\u5b66\u4e60\u8d1d\u53f6\u65af\u540e\u9a8c<\/span>\uff08posterior\uff09<\/span>\u6765\u4fdd\u6301\u5bf9\u4e8e\u4e16\u754c\u6a21\u578b\u7684\u4e0d\u786e\u5b9a\u6027\uff0c\u5e76\u5728\u51b3\u7b56\u7684\u65f6\u5019\u9700\u8981\u8003\u8651\u5230\u8fd9\u4e9b\u4e0d\u786e\u5b9a\u6027\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/p>\n
\u901a\u8fc7\u5b66\u4e60\u8d1d\u53f6\u65af\u540e\u9a8c\u4fdd\u8bc1\u5b89\u5168<\/span><\/section>\n

<\/span><\/section>\n<\/figure>\n
\u4e3e\u4e00\u4e2a\u7b80\u5355\u7684\u4f8b\u5b50\uff1a\u5047\u8bbe\u6211\u4eec\u6709\u4e24\u9053\u95e8\u548c\u89c2\u5bdf\u5230\u7684\u6570\u636e\uff0c\u5b58\u5728\u4e24\u79cd\u201c\u7406\u8bba\u201d\u80fd\u591f\u89e3\u91ca\u8fd9\u7ec4\u6570\u636e\uff1a\u7406\u8bbaA\u662f\u5de6\u95e8\u6709\u6bd2\u53f3\u95e8\u6709\u5956\uff1b\u7406\u8bbaB\u662f\u5de6\u8fb9\u6709\u5956\u53f3\u8fb9\u4e3a\u7a7a\u3002\u6700\u5927\u4f3c\u7136\u7684\u65b9\u6cd5\u53ea\u4f1a\u5b66\u5230\u5176\u4e2d\u4e00\u79cd\u7406\u8bba<\/span>\uff08\u6bd4\u5982\u4e00\u5957\u795e\u7ecf\u7f51\u7edc\u53c2\u6570\u8868\u793a\u7684\u6a21\u578b\uff09<\/span>\uff0c\u56e0\u6b64\u53ef\u80fd\u67091\/4\u7684\u6982\u7387\u62bd\u5230\u6bd2\u836f\u3002\u800c\u8d1d\u53f6\u65af\u6a21\u578b\u9700\u8981\u4fdd\u6301\u5bf9\u4e24\u79cd\u7406\u8bba\u7684\u4e0d\u786e\u5b9a\u6027\uff0c\u56e0\u6b64\u6700\u4f18\u7684\u9009\u62e9\u662f\u5f00\u53f3\u95e8\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/p>\n

Provable Safety for AI Agent: Rejecting Unsafe Actions with Quantitative Guarantee<\/span><\/p>\n


<\/span><\/section>\n<\/figure>\n
\u7ed9\u5b9a\u89c2\u5bdf\u5230\u7684\u6570\u636eD<\/span>\u3001\u4e0a\u4e0b\u6587 x<\/span><\/span>\uff0c\u5047\u8bbe\u6211\u4eec\u80fd\u591f\u5b66\u4e60\u5230\u4e00\u4e2a\u597d\u7684\u8d1d\u53f6\u65af\u540e\u9a8c<\/span>\uff0c\u4e00\u4e2a\u7b80\u5355\u7684\u4fdd\u8bc1\u5b89\u5168\u7684\u65b9\u6cd5\u662f\u6211\u4eec\u53ef\u4ee5\u5728\u5f53 <\/span><\/span> \u5927\u4e8e\u5b89\u5168\u9608\u503c\u7684\u65f6\u5019\u62d2\u7edd\u8fd9\u4e2a\u884c\u52a8\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/span><\/section>\n
<\/span><\/section>\n
<\/span><\/section>\n
<\/span><\/section>\n
<\/p>\n
\u5b66\u4e60\u8d1d\u53f6\u65af\u540e\u9a8c\u7684\u6311\u6218<\/span><\/section>\n

<\/span><\/section>\n<\/figure>\n
\u5f53\u7136\uff0c\u5982\u4f55\u5b66\u5230\u8fd9\u4e2a\u540e\u9a8c\u672c\u8eab\u662f\u4e00\u4e2a\u6311\u6218\uff0c\u7531\u4e8e\u201c\u7406\u8bba\u201d\uff08\u03b8<\/span>\uff09\u7684\u6570\u91cf\u662f\u6307\u6570\u7684\uff0c\u56e0\u6b64\u5728\u91c7\u6837<\/span>\uff08sampling\uff09<\/span>\u548c\u8fb9\u7f18\u5316<\/span>\uff08marginalizing\uff09<\/span>\u90fd\u4f1a\u9047\u5230\u4e0d\u53ef\u76f4\u63a5\u8ba1\u7b97<\/span>\uff08intractable\uff09<\/span>\u7684\u95ee\u9898\u3002<\/span><\/section>\n
<\/p>\n
Bengio\u63d0\u51fa\u7684\u8303\u5f0f\uff1a\u66f4\u591a\u7684\u7b97\u529b\u5f97\u5230\u66f4\u9ad8\u7684\u5b89\u5168\u4fdd\u8bc1<\/span><\/section>\n

<\/span><\/section>\n<\/figure>\n
Bengio\u8ba4\u4e3a\u6b63\u786e\u7684\u8303\u5f0f\u5e94\u8be5\u662f\u80fd\u591f\u901a\u8fc7\u5927\u6a21\u578b\u8fdb\u884c\u644a\u9500\u63a8\u65ad<\/span>\uff08amortized inference\uff09<\/span>\u903c\u8fd1\u540e\u9a8c\uff0c\u5e76\u4e14\u80fd\u591f\u5728\u7b97\u529b\u589e\u52a0\u7684\u8fc7\u7a0b\u5f53\u4e2d\u66f4\u597d\u5730\u5b66\u4e60\u5230\u8d1d\u53f6\u65af\u540e\u9a8c\u7684\u65b9\u6cd5\u3002\u8fd9\u5c31\u542f\u793a\u4e86Bengio\u56e2\u961f\u81ea\u5df1\u7684\u5de5\u4f5c\u751f\u6210\u6d41\u7f51\u7edc<\/span>\uff08GFlowNet\uff09<\/span>\u7684\u65b9\u6cd5\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/p>\n
GFlowNets \u548c \u8d1d\u53f6\u65af\u540e\u9a8c<\/span><\/section>\n

<\/span><\/section>\n<\/figure>\n
\u66f4\u591a\u5185\u5bb9\u53ef\u4ee5\u53c2\u8003Bengio\u7684\u6f14\u8bb2\u89c6\u9891[1]\u3002<\/span><\/section>\n

<\/span><\/section>\n


<\/p>\n

\n
\n
<\/section>\n<\/section>\n<\/section>\n

PART 2 Adam Gleave – AGI Safety: Risks and Research Directions[2]<\/strong><\/span><\/h1>\n

<\/span><\/section>\n
<\/p>\n
AGI Safety: Risks and Research Directions<\/span><\/section>\n

<\/span><\/section>\n<\/figure>\n
\u8bb2\u8005\u4ecb\u7ecd\uff1aAdam Gleave \u662f\u975e\u76c8\u5229\u7814\u7a76\u673a\u6784 FAR AI \u7684\u8054\u5408\u521b\u59cb\u4eba\u548c CEO\uff0cFAR AI\u56e2\u961f\u4e3b\u8981\u5173\u6ce8\u7684\u7814\u7a76\u65b9\u5411\u6709\uff1a\u5bf9\u6297\u9c81\u68d2\u6027\u3001\u4ef7\u503c\u5bf9\u9f50\u548c\u5927\u6a21\u578b\u8bc4\u6d4b\u7b49\u3002Adam Gleave \u5728 UC Berkeley\u7684\u4eba\u7c7b\u517c\u5bb9\u4eba\u5de5\u667a\u80fd\u4e2d\u5fc3<\/span>\uff08Center for Human-Compatible AI\uff09<\/span>\u5b8c\u6210 AI \u535a\u58eb\u5b66\u4f4d\uff0c\u4ed6\u7684\u5bfc\u5e08\u662f\u8457\u540d\u7684 AI \u7814\u7a76\u8005\u65af\u56fe\u5c14\u7279\u00b7\u7f57\u7d20<\/span>\uff08Stuart Russell\uff09<\/span>\u3002Adam\u66fe\u5728 Google DeepMind \u4e0e Jan Leike \u548c Geoffrey Irving \u5408\u4f5c\u8fdb\u884cAI\u5bf9\u9f50\u65b9\u9762\u7684\u7814\u7a76\uff0c\u5e76\u66fe\u5728\u5251\u6865\u5927\u5b66\u4e0e Zoubin Ghahramani \u548c Christian Steinruecken \u5408\u4f5c\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u5173\u4e8e\u4ec0\u4e48\uff1a\u6f14\u8bb2\u4ecb\u7ecd\u4e86AGI\u5b89\u5168\u7684\u5386\u53f2\u6cbf\u9769\u548c\u98ce\u9669\u7814\u5224\uff0c\u5e76\u5c06\u6574\u4e2a\u5de5\u4f5c\u574a\u7684\u4e3b\u9898\u5206\u6210\u4e86\u76d1\u7763<\/span>\uff08Oversight\uff09<\/span>\u3001\u9c81\u68d2\u6027<\/span>\uff08Robustness\uff09<\/span>\u3001\u53ef\u89e3\u91ca\u6027<\/span>\uff08Interpretability\uff09<\/span>\u3001\u6cbb\u7406<\/span>\uff08Governance\uff09<\/span>\u56db\u7c7b\u7814\u7a76\u65b9\u5411\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u5bf9\u4e8eAI\u98ce\u9669\u7684\u5224\u65ad\uff0cGleave\u5f15\u7528\u4e86Sharkey et al. 2023\u5728\u300aA Causal Framework for AI Regulation and Auditing\u300b\u4e2d\u5bf9\u4e8eAI\u98ce\u9669\u7684\u63cf\u8ff0\uff0c\u5f3a\u8c03\u4e86\u5f53AI\u7cfb\u7edf\u62e5\u6709\u66f4\u5927\u7684\u80fd\u529b\u548c\u66f4\u5e7f\u6cdb\u7684\u6388\u6743<\/span>\uff08affordance\uff09<\/span>\u65f6\uff0c\u98ce\u9669\u4f1a\u589e\u52a0\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/p>\n
\u5f53AI\u7cfb\u7edf\u62e5\u6709\u66f4\u5927\u7684\u80fd\u529b\u548c\u66f4\u5e7f\u6cdb\u7684\u6388\u6743\uff08affordance\uff09\u65f6\uff0c\u98ce\u9669\u4f1a\u589e\u52a0<\/span><\/section>\n

<\/span><\/section>\n<\/figure>\n
Gleave\u5f3a\u8c03\u4e86\u4e24\u79cd\u5927\u89c4\u6a21\u98ce\u9669\uff1a\u6ee5\u7528\u98ce\u9669<\/span>\uff08\u6076\u610f\u6216\u9c81\u83bd\u4f7f\u7528AI\u9020\u6210\u4f24\u5bb3\uff09<\/span>\u4e0e\u5931\u63a7AI<\/span>\uff08AI\u81ea\u4e3b\u8ffd\u6c42\u672a\u5bf9\u9f50\u7684\u76ee\u6807<\/span>\uff09\u5e76\u5217\u51fa\u4e86AGI\u5b89\u5168\u7684\u56db\u4e2a\u7814\u7a76\u65b9\u5411\uff1a\u76d1\u7763<\/span>\uff08Oversight\uff09<\/span>\u3001\u9c81\u68d2\u6027<\/span>\uff08Robustness\uff09<\/span>\u3001\u53ef\u89e3\u91ca\u6027<\/span>\uff08Interpretability\uff09<\/span>\u3001\u6cbb\u7406<\/span>\uff08Governance\uff09<\/span>\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/p>\n
\u4e24\u79cd\u5927\u89c4\u6a21\u98ce\u9669\uff1a\u6ee5\u7528\u98ce\u9669\u548c\u5931\u63a7AI<\/span><\/section>\n

<\/span><\/section>\n<\/figure>\n
<\/p>\n
AGI\u5b89\u5168\u7684\u56db\u4e2a\u7814\u7a76\u65b9\u5411<\/span><\/section>\n

<\/span><\/section>\n<\/figure>\n
    \n
  1. \n

    \u76d1\u7763<\/span>\uff08Oversight\uff09<\/span>\u7814\u7a76\u8bd5\u56fe\u660e\u786e\u7cfb\u7edf\u5e94\u5982\u4f55\u884c\u52a8\u4ee5\u6ee1\u8db3\u8bbe\u8ba1\u8005\u8bbe\u5b9a\u7684\u76ee\u6807\u3002<\/span><\/p>\n<\/li>\n<\/ol>\n

      \n
    1. \n

      \u9c81\u68d2\u6027<\/span>\uff08Robustness\uff09<\/span>\u7814\u7a76\u8bd5\u56fe\u786e\u4fdd\u7cfb\u7edf\u5728\u4e00\u7cfb\u5217\u8bbe\u7f6e\u4e2d\u80fd\u591f\u53ef\u9760\u5730\u6ee1\u8db3\u8bbe\u8ba1\u89c4\u8303\u3002<\/span><\/p>\n<\/li>\n

    2. \n

      \u53ef\u89e3\u91ca\u6027<\/span>\uff08Interpretability\uff09<\/span>\u7814\u7a76\u8bd5\u56fe\u4f7f\u4eba\u7c7b\u7406\u89e3\u7cfb\u7edf\u7684\u80fd\u529b\u548c\u5c40\u9650\u6027\u3002<\/span><\/p>\n<\/li>\n

    3. \n
      \u6cbb\u7406<\/span>\uff08Governance\uff09<\/span>\u7814\u7a76\u8bd5\u56fe\u534f\u8c03\u7cfb\u7edf\u8bad\u7ec3\u548c\u90e8\u7f72\u7684\u5b89\u5168\u6807\u51c6\u3002<\/span><\/section>\n<\/li>\n<\/ol>\n


      <\/h1>\n


      <\/p>\n

      \n
      \n
      <\/section>\n<\/section>\n<\/section>\n

      PART 3 Owain Evans – Out-of-context Reasoning in LLMs[3]<\/span><\/strong><\/span><\/h1>\n


      <\/span><\/strong><\/span><\/p>\n

      <\/p>\n
      Out-of-context Reasoning in LLMs (Owain Evans)<\/span><\/section>\n

      <\/span><\/section>\n<\/figure>\n
      \u8bb2\u8005\u4ecb\u7ecd\uff1aOwain Evans\u76ee\u524d\u6b63\u5728\u9886\u5bfc\u4e00\u4e2a\u4f4d\u4e8e\u7f8e\u56fd\u52a0\u5dde\u4f2f\u514b\u5229\u7684AI\u5b89\u5168\u7814\u7a76\u7ec4\u3002\u4ed6\u7684\u7814\u7a76\u5174\u8da3\u5305\u62ecAI\u5bf9\u9f50\u548cAGI\u98ce\u9669\u3002\u4ed6\u76ee\u524d\u5173\u6ce8\u5927\u6a21\u578b\u7684\u6001\u52bf\u611f\u77e5<\/span>\uff08situational awareness\uff09<\/span>\u4e0e\u6b3a\u9a97\u80fd\u529b\uff0c\u4ee5\u53caAI\u7cfb\u7edf\u7684\u771f\u5b9e\u6027\u548c\u8bda\u5b9e\u6027\u3002\u4ed6\u66fe\u5728\u725b\u6d25\u5927\u5b66\u7684\u672a\u6765\u4eba\u7c7b\u7814\u7a76\u6240\u8fdb\u884cAI\u5bf9\u9f50\u7814\u7a76\uff0c\u5e76\u5728MIT\u83b7\u5f97PhD\u5b66\u4f4d\u3002\u4ed6\u5173\u4e8eAI\u5b89\u5168\u7684\u5de5\u4f5c\u5305\u62ec\uff1a\u9006\u8f6c\u8bc5\u5492<\/span>\uff08Reversal Curse\uff09<\/span>[4]<\/span>\uff08<\/span>\u4e2d\u6587\u89e3\u8bfb<\/span><\/a>\uff09<\/span>\u3001TruthfulQA\u7b49\u3002<\/span><\/section>\n

      <\/span><\/section>\n
      \u5173\u4e8e\u4ec0\u4e48\uff1aEvans\u5728\u6f14\u8bb2\u4e2d\u7814\u7a76\u7684\u6838\u5fc3\u6982\u5ff5\u53eb\u505a\u201c\u4e0a\u4e0b\u6587\u5916\u63a8\u7406\u201d<\/span>\uff08Out-of-context Reasoning\uff1bOOCR\uff09<\/span>\u3002\u8fd9\u79cd\u80fd\u529b\u6307\u7684\u662f\u201c\u5728\u4e0d\u4f7f\u7528\u601d\u7ef4\u94fe<\/span>\uff08Chain-of-Thought\uff09<\/span>\u7684\u60c5\u51b5\u4e0b\u8fdb\u884c\u903b\u8f91\u63a8\u7406\u5e76\u5f97\u51fa\u7ed3\u8bba\u7684\u80fd\u529b\u201d\uff0c\u8fd9\u79cd\u63a8\u7406\u65b9\u5f0f\u4e0e\u201c\u4e0a\u4e0b\u6587\u5185\u63a8\u7406\u201d<\/span>\uff08in-context\uff09<\/span>\u76f8\u5bf9\u3002<\/span><\/section>\n

      <\/span><\/section>\n
      <\/p>\n
      \u201cin-context learning\u201d \u4e0e \u201cout-of-context learning\u201d \u7684\u5bf9\u6bd4<\/span><\/section>\n

      <\/span><\/section>\n<\/figure>\n
      \u4e00\u822c\u6765\u8bf4\u6211\u4eec\u4f1a\u4ee5\u201c\u4e0a\u4e0b\u6587\u5185\u63a8\u7406\u201d<\/span>\uff08in-context\uff09<\/span>\u7684\u5f62\u5f0f\u4f7f\u7528LLM\u3002\u6211\u4eec\u628a\u4e00\u7cfb\u5217\u524d\u63d0\u6761\u4ef6<\/span>\uff08premises\uff09<\/span>\u653e\u5728\u4e0a\u4e0b\u6587\u7a97\u53e3\u5185\uff0c\u5e76\u53ef\u80fd\u4f7f\u7528\u601d\u7ef4\u94fe\u63d0\u793a\u6a21\u578b\u5c06\u591a\u6b65\u63a8\u7406\u624d\u80fd\u89e3\u51b3\u7684\u95ee\u9898\u5206\u89e3\u6210\u5355\u6b65\u63a8\u7406\uff0c\u4ee5\u81ea\u7136\u8bed\u8a00\u7684\u5f62\u5f0f\u663e\u5f0f\u5730\u4e00\u6b65\u6b65\u5f97\u5230\u7ed3\u8bba\u3002\u4f46OOCR\u60f3\u8981\u7814\u7a76\u7684\u662f\u628a\u8fd9\u4e9b\u524d\u63d0\u6761\u4ef6\u653e\u5728\u591a\u4e2a\u8bad\u7ec3\u6570\u636e\u4e2d\uff0c\u662f\u5426\u80fd\u5728\u4e0d\u4f7f\u7528\u601d\u7ef4\u94fe\u7684\u60c5\u51b5\u4e0b\u76f4\u63a5\u5f97\u5230\u7ed3\u8bba\uff1f<\/span><\/section>\n

      <\/span><\/section>\n
      \u5b9e\u8bc1\u7ed3\u679c\uff1a\u56e2\u961f\u4e8e2023\u5e749\u6708\u53d1\u5e03\u7684\u6587\u7ae0\u300aTaken out of context: On measuring situational awareness in LLMs\u300b[5]\u8bbe\u8ba1\u4e86\u5b9e\u9a8c\u6d4b\u8bd5\u4e86\u5f53\u524d\u6a21\u578b\u7684\u521d\u59cbOOCR\u80fd\u529b\u3002<\/span><\/section>\n

      <\/span><\/section>\n
      <\/p>\n
      \u300aTaken out of context: On measuring situational awareness in LLMs\u300b<\/span><\/section>\n

      <\/span><\/section>\n<\/figure>\n
      \u53e6\u5916\u4f5c\u8005\u4e5f\u505a\u4e86\u62d3\u5c55\u5b9e\u9a8c\uff0c\u53d1\u73b0\u66f4\u5927\u7684\u6a21\u578b\u503e\u5411\u4e8e\u5728\u8fd9\u4e9b\u9700\u8981 OOCR \u624d\u80fd\u5b8c\u6210\u7684\u4efb\u52a1\u4e0a\u8868\u73b0\u5f97\u66f4\u597d\uff0c\u4f46\u662f\u603b\u4f53\u7684\u63d0\u5347\u76f8\u5bf9\u4e8e\u5176\u4ed6\u6d8c\u73b0\u7684\u80fd\u529b\u8fd8\u4e0d\u8db3\u3002\u591a\u6b65\u63a8\u7406\u76f8\u5bf9\u4e8e\u5355\u6b65\u63a8\u7406\u66f4\u52a0\u56f0\u96be\u3002<\/span><\/section>\n

      <\/span><\/section>\n
      <\/p>\n
      OOCR\u80fd\u529b\u7684\u62d3\u5c55\u5b9e\u9a8c<\/span><\/section>\n

      <\/span><\/section>\n<\/figure>\n
      \u56e2\u961f\u8fd8\u53d1\u73b0\u4e00\u4e9b\u7b80\u5355\u7684OOCR\u80fd\u529b\u5904\u4e8e\u5b8c\u5168\u5931\u8d25\u7684\u72b6\u6001\u3002\u4e00\u4e2a\u7ecf\u5178\u7684\u4f8b\u5b50\u662f\u9006\u8f6c\u8bc5\u5492<\/span>\uff08Reversal Curse\uff09\uff08<\/span>\u4e2d\u6587\u89e3\u8bfb<\/span><\/a>\uff09<\/span>\u3002\u7b80\u5355\u6765\u8bf4\uff0c\u4f5c\u8005\u628aLLM\u5728\u751f\u6210\u7684\u201cA\u662fB\u201d\u7684\u4e8b\u5b9e\u6570\u636e\u4e0a\u5fae\u8c03\uff0c\u7136\u540e\u95eeLLM\u201cB\u662f\u4ec0\u4e48\uff1f\u201d\uff0c\u4f5c\u8005\u53d1\u73b0LLM\u56de\u7b54\u4e0d\u51fa\u6765\u3002<\/span><\/section>\n

      <\/span><\/section>\n
      <\/p>\n
      \u9006\u8f6c\u8bc5\u5492\uff08Reversal Curse\uff09<\/span><\/section>\n

      <\/span><\/section>\n<\/figure>\n
      \u56e0\u6b64\u4f5c\u8005\u53d1\u73b0\uff1a\u81ea\u56de\u5f52\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b<\/span>\uff08LLMs\uff09<\/span>\u4e0d\u80fd\u8fdb\u884c\u4efb\u4f55\u4f9d\u8d56\u4e8e\u53cd\u8f6c\u524d\u63d0\u987a\u5e8f\u7684OOCR\u3002\u5373\u4f7f\u5c1d\u8bd5\u4e86\u6269\u5927\u6a21\u578b\u89c4\u6a21\u3001\u6570\u636e\u589e\u5f3a\u548c\u5176\u4ed6\u65b9\u6cd5\uff0c\u90fd\u6ca1\u6709\u4efb\u4f55\u5e2e\u52a9\u3002<\/span><\/section>\n

      <\/span><\/section>\n
      <\/p>\n
      OOC Reasoning \u4e3a\u4f55\u91cd\u8981\uff1f<\/span><\/section>\n<\/figure>\n

      <\/section>\n
      \u4e3a\u4f55\u91cd\u8981\uff1a\u5047\u8bbe\u6a21\u578b\u7684 OOCR \u80fd\u529b\u5f88\u5f3a\uff0c\u5e76\u4e14\u968f\u7740\u6a21\u578b\u89c4\u6a21\u589e\u5927\u800c\u53d8\u5f3a\uff0c\u90a3\u4e48\u672a\u6765\u7684\u6a21\u578b\u53ef\u80fd\u4f1a\u5728\u4e0d\u4f7f\u7528\u601d\u7ef4\u94fe\u7684\u60c5\u51b5\u4e0b\uff0c\u79ef\u7d2f\u65b0\u77e5\u8bc6\u548c\u65b0\u7ed3\u8bba\uff1b\u6a21\u578b\u53ef\u80fd\u4f1a\u79ef\u7d2f\u9690\u85cf\u7684\u7b56\u7565\u548c\u89c4\u5212<\/span>\uff08\u6bd4\u5982\u64cd\u63a7\u548c\u6b3a\u9a97\u4eba\u7c7b\uff09<\/span>\uff1b\u6a21\u578b\u4e5f\u53ef\u80fd\u4f1a\u731c\u6d4b\u51fa\u4eba\u7c7b\u4ece\u8bad\u7ec3\u6570\u636e\u4e2d\u6392\u9664\u7684\u4e8b\u5b9e<\/span>\uff08\u4f8b\u5982\uff1a\u5982\u4f55\u5236\u9020\u6b66\u5668\u4ee5\u53ca\u5173\u4e8e\u6a21\u578b\u672c\u8eab\u7684\u77e5\u8bc6\uff09<\/span>\u3002\u76f8\u53cd\u7684\uff0c\u5982\u679c\u6a21\u578b\u7684 OOCR \u80fd\u529b\u4e0d\u8db3\u4e14\u96be\u4ee5\u62d3\u5c55\uff0c\u90a3\u4e48 LLM \u7684\u77e5\u8bc6\u53ea\u4f1a\u5c40\u9650\u4e8e\u4eba\u7c7b\u7684\u77e5\u8bc6\u3001\u81ea\u84b8\u998f\u548c\u601d\u7ef4\u94fe\u8bbe\u7f6e\u7684\u63a8\u7406\u8fb9\u754c\u3002\u4e3a\u4e86\u4fdd\u8bc1\u5b89\u5168\uff0c\u6211\u4eec\u5c31\u4e5f\u53ef\u4ee5\u7528\u5176\u4ed6\u6a21\u578b\u76d1\u6d4b\u601d\u7ef4\u94fe\u7684\u6b65\u9aa4\u6765\u963b\u6b62\u53ef\u7591\u7684\u6b65\u9aa4\u3002<\/span><\/section>\n

      <\/span><\/section>\n

      <\/section>\n
      \u53c2\u8003\u8d44\u6599<\/span><\/strong><\/section>\n
      [1] <\/span>Yoshua Bengio: Towards Quantitative Safety Guarantees and Alignment: https:\/\/www.alignment-workshop.com\/nola-2023<\/span><\/section>\n
      [2] <\/span>Adam Gleave – AGI Safety: Risks and Research Directions: https:\/\/w<\/span>ww.alignment-workshop.com\/nola-talks\/adam-gleave-agi-safety-risks-and-research-directions<\/span><\/section>\n
      [3] <\/span>Owain Evans – Out-of-context Reasoning in LLMs: https:\/\/www.alignment-workshop.com\/nola-talks\/owain-evans-out-of-context-reasoning-in-llms<\/span><\/section>\n
      [4] <\/span>\u9006\u8f6c\u8bc5\u5492\uff08Reversal Curse\uff09: https:\/\/arxiv.org\/abs\/2309.12288<\/span><\/section>\n
      [5] <\/span>\u300aTaken out of context: On measuring situational awareness in LLMs\u300b: https:\/\/arxiv.org\/abs\/2309.00667<\/span><\/section>\n
      \n

      <\/section>\n

      <\/section>\n
      \n
      \n
      \n
      \n

      \u6d3b\u52a8\u9884\u544a<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n


      <\/p>\n

      <\/mp-common-videosnap><\/section>\n
      <\/section>\n


      <\/p>\n

      <\/span><\/section>\n

      \u5927\u6a21\u578b\u7684\u72c2\u98d9\u7a81\u8fdb\u5524\u9192\u4e86\u4eba\u4eec\u5bf9AI\u6280\u672f\u7684\u70ed\u60c5\u548c\u61a7\u61ac\uff0c\u4e5f\u5f15\u53d1\u4e86\u5bf9AI\u6280\u672f\u672c\u8eab\u5b58\u5728\u7684\u793e\u4f1a\u4f26\u7406\u98ce\u9669\u53ca\u5176\u5bf9\u4eba\u7c7b\u751f\u5b58\u6784\u6210\u7684\u6f5c\u5728\u5a01\u80c1\u7684\u666e\u904d\u62c5\u5fe7\u3002\u5728\u6b64\u80cc\u666f\u4e0b\uff0cAI\u5b89\u5168\u4e0e\u5bf9\u9f50\u5f97\u5230\u5e7f\u6cdb\u5173\u6ce8\uff0c\u8fd9\u662f\u4e00\u4e2a\u81f4\u529b\u4e8e\u8ba9AI\u9020\u798f\u4eba\u7c7b\uff0c\u907f\u514dAI\u6a21\u578b\u5931\u63a7\u6216\u88ab\u6ee5\u7528\u800c\u5bfc\u81f4\u707e\u96be\u6027\u540e\u679c\u7684\u7814\u7a76\u65b9\u5411\u3002\u96c6\u667a\u4ff1\u4e50\u90e8\u548c\u5b89\u8fdcAI\u8054\u5408\u4e3e\u529e\u300cAI\u5b89\u5168\u4e0e\u5bf9\u9f50\u300d\u8bfb\u4e66\u4f1a<\/a>\uff0c\u7531\u591a\u4f4d\u6d77\u5185\u5916\u4e00\u7ebf\u7814\u7a76\u8005\u8054\u5408\u53d1\u8d77\uff0c\u65e8\u5728\u6df1\u5165\u63a2\u8ba8AI\u5b89\u5168\u4e0e\u5bf9\u9f50\u6240\u6d89\u53ca\u7684\u6838\u5fc3\u6280\u672f\u3001\u7406\u8bba\u67b6\u6784\u3001\u89e3\u51b3\u8def\u5f84\u4ee5\u53ca\u5b89\u5168\u6cbb\u7406\u7b49\u4ea4\u53c9\u8bfe\u9898\u3002<\/span><\/p>\n


      <\/strong><\/span><\/section>\n
      <\/span><\/a><\/section>\n

      <\/section>\n

      <\/section>\n
      \u8be6\u60c5\u8bf7\u89c1\uff1a<\/span><\/section>\n
      2024\u5f00\u5e74\u8bfb\u4e66\u4f1a\uff1aAI\u5b89\u5168\u4e0e\u5bf9\u9f50\u2014\u2014\u5e94\u5bf9\u524d\u6cbfAI\u5931\u63a7\u4e0e\u6ee5\u7528\u7684\u6280\u672f\u8def\u7ebf<\/a><\/span>
      <\/span><\/section>\n

      <\/section>\n<\/section>\n
      \n
      \n
      \n
      \n
      \n

      \u63a8\u8350\u9605\u8bfb<\/span><\/strong><\/strong><\/p>\n

      1.  <\/strong>\u5927\u6a21\u578b\u5b89\u5168\u4e0e\u5bf9\u9f50\uff1a\u590d\u6742\u7cfb\u7edf\u89c6\u89d2\u4e0b\u7684AI\u5b89\u5168<\/a><\/strong><\/span><\/strong><\/strong><\/section>\n
      2. <\/strong>\u4e07\u5b57\u957f\u6587\u8be6\u89e3\uff1a\u5927\u6a21\u578b\u65f6\u4ee3AI\u4ef7\u503c\u5bf9\u9f50\u7684\u95ee\u9898\u3001\u5bf9\u7b56\u548c\u5c55\u671b<\/a><\/strong><\/span><\/strong><\/section>\n
      3. <\/strong>AI\u4f55\u4ee5\u6d8c\u73b0\uff1a\u590d\u6742\u9002\u5e94\u7cfb\u7edf\u89c6\u89d2\u7684ChatGPT\u548c\u5927\u8bed\u8a00\u6a21\u578b<\/a><\/strong><\/span><\/strong><\/section>\n
      4. <\/strong><\/span>\u5f20\u6c5f\uff1a\u7b2c\u4e09\u4ee3\u4eba\u5de5\u667a\u80fd\u6280\u672f\u57fa\u7840\u2014\u2014\u4ece\u53ef\u5fae\u5206\u7f16\u7a0b\u5230\u56e0\u679c\u63a8\u7406 | \u96c6\u667a\u5b66\u56ed\u5168\u65b0\u8bfe\u7a0b<\/a><\/strong><\/strong><\/span><\/section>\n
      5. <\/strong>\u52a0\u5165\u96c6\u667a\u5b66\u56edVIP\uff0c\u4e00\u6b21\u6027\u83b7\u53d6\u96c6\u667a\u5e73\u53f0\u6240\u6709\u5185\u5bb9\u8d44\u6e90<\/a><\/strong><\/strong><\/strong><\/strong><\/span><\/section>\n
      6. <\/strong>\u52a0\u5165\u96c6\u667a\uff0c\u4e00\u8d77\u590d\u6742\uff01<\/strong><\/span><\/a><\/section>\n<\/section>\n


      <\/p>\n<\/section>\n


      <\/p>\n

      \u70b9\u51fb\u201c\u9605\u8bfb\u539f\u6587\u201d\uff0c\u62a5\u540d\u8bfb\u4e66\u4f1a<\/span><\/strong><\/section>\n<\/section>\n<\/section>\n
      \n
      \n

      <\/section>\n<\/section>\n<\/section>\n<\/section>\n

      <\/mp-style-type><\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

      \u5bfc\u8bed \u5728 2023 \u5e74 12 \u6708\uff0c\u7b2c\u4e8c\u5c4a AI \u5bf9\u9f50\u5de5\u4f5c\u574a\uff08Alignment Workshop\uff09\u5728\u65b0\u5965\u5c14\u826f\u7684 NeurIPS \u4f1a\u8bae\u671f\u95f4\u4e3e\u529e\u3002\u672c\u6b21\u5de5\u4f5c\u574a\u662f\u7531 Adam Gleave \u9886\u5bfc\u7684\u975e\u76c8\u5229\u6027\u7814\u7a76\u673a\u6784 FAR AI \u4e3b\u529e\u3002 \u5de5\u4f5c\u574a\u9080\u8bf7\u5230\u4e86\u6765\u81ea\u5de5\u4e1a\u754c\u548c\u5b66\u672f\u754c150\u4f59\u4f4dAI\u7814\u7a76\u8005\u5c31 AI \u5b89\u5168\u548c\u5bf9\u9f50…<\/p>\n","protected":false},"author":0,"featured_media":47992,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[1],"tags":[],"special":[],"_links":{"self":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts\/48027"}],"collection":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=48027"}],"version-history":[{"count":0,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts\/48027\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/media\/47992"}],"wp:attachment":[{"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=48027"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=48027"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=48027"},{"taxonomy":"special","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fspecial&post=48027"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}