{"id":48796,"date":"2024-03-07T20:24:09","date_gmt":"2024-03-07T12:24:09","guid":{"rendered":"https:\/\/swarma.org\/?p=48796"},"modified":"2024-03-07T20:24:09","modified_gmt":"2024-03-07T12:24:09","slug":"agi%e6%80%8e%e6%a0%b7%e5%8f%8d%e8%af%88%ef%bc%9a%e5%af%b9%e6%8a%97%e9%b2%81%e6%a3%92%e6%80%a7%e4%b8%8e%e5%af%b9%e9%bd%90%e6%b3%9b%e5%8c%96%e8%a7%86%e8%a7%92-%e5%a4%a7%e6%a8%a1","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=48796","title":{"rendered":"AGI\u600e\u6837\u201c\u53cd\u8bc8\u201d\uff1a\u5bf9\u6297\u9c81\u68d2\u6027\u4e0e\u5bf9\u9f50\u6cdb\u5316\u89c6\u89d2 | \u5927\u6a21\u578b\u5b89\u5168\u4e0e\u5bf9\u9f50\u8bfb\u4e66\u4f1a\u00b7\u5468\u516d\u76f4\u64ad"},"content":{"rendered":"
\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
\u5bf9\u6297\u6837\u672c\u7684\u5b58\u5728\u5bf9\u901a\u7528\u4eba\u5de5\u667a\u80fd\u5728\u73b0\u5b9e\u573a\u666f\u4e2d\u7684\u5b89\u5168\u5e94\u7528\u6784\u6210\u4e86\u4e25\u91cd\u7684\u5a01\u80c1\u3002\u5982\u4f55\u4ece\u5bf9\u6297\u9c81\u68d2\u6027\u7684\u89c6\u89d2\u589e\u5f3a\u901a\u7528\u4eba\u5de5\u667a\u80fd\u62b5\u5fa1\u653b\u51fb\u7684\u80fd\u529b\uff1f\u76ee\u524d\u6709\u54ea\u4e9b\u4fc3\u8fdb\u5bf9\u6297\u9c81\u68d2\u6027\u7684\u6700\u4f73\u65b9\u6cd5\uff1f\u8fd9\u4e9b\u65b9\u6cd5\u5b58\u5728\u54ea\u4e9b\u4e0d\u8db3\uff1f\u800c\u5728\u4ef7\u503c\u504f\u597d\u5bf9\u9f50\u9886\u57df\uff0c\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09\u548c\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08DPO\uff09\u4f5c\u4e3a\u4e3b\u8981\u6280\u672f\uff0c\u5e7f\u6cdb\u5e94\u7528\u4e8e\u5148\u8fdb\u7684\u5927\u578b\u6a21\u578b\u3002\u8fd9\u4e9b\u65b9\u6cd5\u5c5e\u4e8e\u8bad\u7ec3\u65f6\u5bf9\u9f50\uff0c\u6d89\u53ca\u6a21\u578b\u8bad\u7ec3\uff0c\u6210\u672c\u8f83\u9ad8\u3002\u6709\u6ca1\u6709\u66f4\u597d\u7684\u5bf9\u9f50\u65b9\u6cd5\uff1f\u5b83\u5728\u5e94\u5bf9jailbreaking\u7b49\u653b\u51fb\u65f6\u4f1a\u5b58\u5728\u54ea\u4e9b\u6f5c\u5728\u98ce\u9669\uff1f<\/span><\/strong><\/section>\n

<\/span><\/strong><\/section>\n
\u201c\u5927\u6a21\u578b\u5b89\u5168\u4e0e\u5bf9\u9f50\u201d\u8bfb\u4e66\u4f1a<\/a><\/span><\/strong>\u7684\u7b2c\u4e94\u671f\uff0c\u6211\u4eec\u9080\u8bf7\u5230\u9ebb\u7701\u7406\u5de5\u5b66\u9662\u7684Tony Wang\u548c\u4e0a\u6d77\u4eba\u5de5\u667a\u80fd\u5b9e\u9a8c\u5ba4\u7684\u5468\u5c55\u8f89\u3002\u57fa\u4e8e\u4e0a\u8ff0\u95ee\u9898\uff0cTony Wang\u5c06\u7740\u91cd\u8ba8\u8bba\uff081\uff09\u4ee5\u89c6\u89c9\u4e0e\u96f6\u548c\u535a\u5f08\uff08\u5c24\u5176\u662f\u56f4\u68cb\uff09\u4f5c\u4e3a\u4e24\u4e2a\u6848\u4f8b\u7814\u7a76\uff0c\u6982\u8ff0\u5982\u4f55\u5c06\u901a\u7528\u4eba\u5de5\u667a\u80fd\u5b89\u5168\u89c6\u4e3a\u4e00\u4e2a\u5bf9\u6297\u9c81\u68d2\u6027\u95ee\u9898\u4ee5\u53ca\u8be5\u95ee\u9898\u7684\u96be\u70b9\u6240\u5728\u3002\uff082\uff09\u4fc3\u8fdb\u5bf9\u6297\u9c81\u68d2\u6027\u7684\u6700\u4f73\u65b9\u6cd5\u4ee5\u53ca\u8fd9\u4e9b\u65b9\u6cd5\u7684\u4e0d\u8db3\u4e4b\u5904\u3002\uff083\uff09\u63a8\u52a8\u5bf9\u6297\u9c81\u68d2\u6027\u9886\u57df\u5411\u524d\u53d1\u5c55\u7684\u6f5c\u5728\u65b9\u6cd5\u3002\u968f\u540e\uff0c\u5468\u5c55\u8f89\u5c06\u4ecb\u7ecd\u63a8\u7406\u65f6\u5bf9\u9f50\u65b9\u6cd5\uff08Inference-time Alignment\uff09\uff0c\u4e00\u7cfb\u5217\u65e0\u9700\u989d\u5916\u8bad\u7ec3\uff0c\u4ec5\u4f9d\u9760\u8f83\u5f31\u6a21\u578b\u7684\u5f15\u5bfc\u4fe1\u53f7(weak supervisor)\u6765\u63d0\u9ad8\u8f83\u5f3a\u6a21\u578b\u80fd\u529b\u65b9\u6cd5\u3002\u540c\u65f6\u8ba8\u8bba\u63a8\u7406\u65f6\u5bf9\u9f50\u7684\u6f5c\u5728\u98ce\u9669\uff0c\u4ee5\u201c\u53cd\u5411\u7684\u63a8\u7406\u65f6\u5bf9\u9f50\u201d\uff08Inference-time Disalignment\uff09\u4e3a\u4f8b\u6765\u8bba\u8bc1\u5f00\u6e90\u6a21\u578b\u7684\u5b89\u5168\u6027\u5728\u63a8\u7406\u9636\u6bb5\u88ab\u4f4e\u6210\u672c\u653b\u51fb(Jailbreak)\u7684\u98ce\u9669\u3002<\/span><\/strong><\/section>\n
<\/mp-common-videosnap><\/section>\n
<\/span><\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n

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

<\/section>\n

\n
\n
\n
\n
\n

<\/section>\n

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

\u5206\u4eab\u5927\u7eb2<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


<\/p>\n

<\/span><\/h2>\n

AGI\u5b89\u5168\u89c6\u89d2\u4e0b\u7684\u5bf9\u6297\u9c81\u68d2\u6027<\/span><\/strong><\/span><\/h3>\n