{"id":5836,"date":"2025-03-30T17:56:13","date_gmt":"2025-03-30T09:56:13","guid":{"rendered":"https:\/\/nullthought.net\/?p=5836"},"modified":"2025-03-30T18:11:23","modified_gmt":"2025-03-30T10:11:23","slug":"anthropic%e5%9f%ba%e4%ba%8ecircuit-tracing%e6%8f%ad%e7%a4%ballm%e5%86%85%e9%83%a8%e6%9c%ba%e5%88%b6","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=5836","title":{"rendered":"Anthropic\u57fa\u4e8ecircuit tracing\u63ed\u793aLLM\u5185\u90e8\u673a\u5236"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">\u2705 <strong>\u9996\u6b21\u5728\u8d85\u5927\u6a21\u578b\uff08Claude 3.5\uff09\u4e2d\u7cfb\u7edf\u6027\u5b9e\u65bdcircuit tracing\u5206\u6790<\/strong>\uff1b<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u2705 \u63ed\u793a\u6a21\u578b\u884c\u4e3a\u4e0d\u4ec5\u57fa\u4e8e\u8f93\u5165\uff0c\u8fd8\u4f9d\u8d56\u590d\u6742\u5185\u90e8\u8def\u5f84\u4e0e\u591a\u7ec4\u4ef6\u6743\u91cd\u8c03\u8282\uff1b<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u2705 \u63d0\u4f9b\u4e86\u7406\u89e3LLM\u884c\u4e3a\u7684\u7ed3\u6784\u5316\u3001\u53ef\u9a8c\u8bc1\u6027\u8def\u5f84\uff1b<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u2705 \u53ef\u80fd\u91cd\u6784\u672a\u6765\u5bf9\u201c<strong><a href=\"https:\/\/nullthought.net\/?p=3965\" target=\"_blank\" rel=\"noreferrer noopener\">\u6a21\u578b\u53ef\u89e3\u91ca\u6027<\/a><\/strong>\u201d\u7684\u5b9a\u4e49\u4e0e\u65b9\u6cd5\u8bba\u57fa\u7840\uff1b<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u2705 \u4e3a\u5f3a\u5316\u5b89\u5168\u63a7\u5236\u3001\u63d0\u5347\u8de8\u8bed\u8a00\u80fd\u529b\u3001\u63d0\u5347\u8bd7\u6027\u6587\u672c\u8d28\u91cf\u63d0\u4f9b\u91cd\u8981\u53c2\u8003\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e00\u3001\u7814\u7a76\u80cc\u666f\uff1a\u63ed\u5f00\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u201c\u9ed1\u7bb1\u201d\u4e4b\u8c1c<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08Large Language Models, LLMs\uff09\u5982GPT\u3001Claude\u7b49\uff0c\u5728\u8fc7\u53bb\u51e0\u5e74\u5c55\u73b0\u4e86\u60ca\u4eba\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u80fd\u529b\uff0c\u80fd\u591f\u751f\u6210\u7c7b\u4f3c\u4eba\u7c7b\u7684\u6587\u672c\u3001\u5199\u4ee3\u7801\u3001\u89e3\u9898\u751a\u81f3\u521b\u4f5c\u8bd7\u6b4c\u3002\u7136\u800c\uff0c\u5b83\u4eec\u7684\u5f3a\u5927\u80fd\u529b\u80cc\u540e\u7684\u8fd0\u4f5c\u673a\u5236\u5374\u9c9c\u6709\u4eba\u771f\u6b63\u7406\u89e3\u3002\u5373\u4f7f\u662f\u6700\u524d\u6cbf\u7684\u7814\u7a76\u4eba\u5458\uff0c\u4e5f\u666e\u904d\u5c06\u8fd9\u4e9b\u6a21\u578b\u79f0\u4e3a\u201c\u9ed1\u7bb1\u201d\uff0c\u56e0\u4e3a\u5b83\u4eec\u7684\u6bcf\u4e00\u4e2a\u8f93\u51fa\u90fd\u6e90\u81ea\u6570\u4ee5\u5341\u4ebf\u8ba1\u7684\u53c2\u6570\uff0c\u4f46\u6211\u4eec\u5f88\u96be\u6e05\u695a\u5730\u77e5\u9053\u8fd9\u4e9b\u53c2\u6570\u662f\u5982\u4f55\u534f\u540c\u5de5\u4f5c\u7684\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u7f3a\u4e4f\u53ef\u89e3\u91ca\u6027\u4e0d\u4ec5\u5e26\u6765\u4e86\u5b66\u672f\u4e0a\u7684\u6311\u6218\uff0c\u4e5f\u5f15\u53d1\u4e86\u5173\u4e8e\u5b89\u5168\u6027\u3001\u63a7\u5236\u6027\u548c\u53ef\u4fe1\u5ea6\u7684\u5e7f\u6cdb\u4e89\u8bba\u3002\u5982\u679c\u6211\u4eec\u65e0\u6cd5\u7406\u89e3\u6a21\u578b\u662f\u5982\u4f55\u505a\u51fa\u51b3\u7b56\u7684\uff0c\u5c31\u96be\u4ee5\u9884\u5224\u5b83\u5728\u590d\u6742\u60c5\u5883\u4e0b\u7684\u884c\u4e3a\u8fb9\u754c\u3001\u9519\u8bef\u98ce\u9669\u4ee5\u53ca\u53ef\u80fd\u7684\u6ee5\u7528\u8def\u5f84\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b63\u662f\u5728\u8fd9\u6837\u7684\u80cc\u666f\u4e0b\uff0cAI\u516c\u53f8Anthropic\u5f00\u5c55\u4e86\u4e00\u9879\u5f00\u521b\u6027\u7814\u7a76\uff0c\u5f00\u53d1\u4e86\u4e00\u79cd\u89c2\u5bdf\u6a21\u578b\u5185\u90e8\u5904\u7406\u8fc7\u7a0b\u7684\u65b0\u6280\u672f\uff0c\u4f7f\u5f97\u7814\u7a76\u8005\u80fd\u591f\u201c\u770b\u89c1\u201dLLM\u5728\u5904\u7406\u63d0\u793a\u65f6\u5185\u90e8\u53d1\u751f\u7684\u4e8b\u60c5\u3002\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u4ec5\u63d0\u4f9b\u4e86\u524d\u6240\u672a\u6709\u7684\u201c\u53ef\u89c6\u5316\u89c6\u89d2\u201d\uff0c\u8fd8\u63ed\u793a\u4e86\u5927\u91cf\u8fdd\u80cc\u76f4\u89c9\u7684\u63a8\u7406\u8def\u5f84\u3001\u7ec4\u4ef6\u6fc0\u6d3b\u673a\u5236\u548c\u884c\u4e3a\u903b\u8f91\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e8c\u3001\u6838\u5fc3\u6280\u672f\u65b9\u6cd5\uff1acircuit tracing\u63ed\u793a\u51b3\u7b56\u8def\u5f84<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b64\u6b21\u7814\u7a76\u7684\u6838\u5fc3\u65b9\u6cd5\u662fcircuit tracing\uff0c\u8fd9\u662f\u4e00\u79cd\u7528\u4e8e\u8ffd\u8e2a\u795e\u7ecf\u7f51\u7edc\u5185\u90e8\u201c\u7ec4\u4ef6\u94fe\u6761\u201d\uff08circuits\uff09\u5728\u5904\u7406\u4efb\u52a1\u65f6\u6fc0\u6d3b\u8def\u5f84\u7684\u6280\u672f\u3002\u5177\u4f53\u800c\u8a00\uff0cLLM\u7531\u65e0\u6570\u4e2a\u5fae\u5c0f\u7684\u201c\u7ec4\u4ef6\u201d\uff08component\uff09\u7ec4\u6210\uff0c\u6bcf\u4e2a\u7ec4\u4ef6\u7531\u591a\u4e2a\u795e\u7ecf\u5143\u6216\u7279\u5b9a\u53c2\u6570\u7ec4\u5408\u6784\u6210\u3002\u8fd9\u4e9b\u7ec4\u4ef6\u53ef\u4ee5\u4ee3\u8868\u67d0\u79cd\u6982\u5ff5\u3001\u8bed\u4e49\u3001\u903b\u8f91\u6216\u72b6\u6001\uff0c\u5982\u989c\u8272\u3001\u60c5\u7eea\u3001\u5730\u540d\u3001\u8bed\u8a00\u7279\u5f81\u7b49\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">circuit tracing\u7684\u5173\u952e\u5728\u4e8e\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u7ec4\u4ef6\u8bc6\u522b<\/strong>\uff1a\u9996\u5148\u786e\u5b9a\u6a21\u578b\u5185\u90e8\u54ea\u4e9b\u7ec4\u4ef6\u5728\u7279\u5b9a\u8bed\u4e49\u6216\u4e0a\u4e0b\u6587\u4e0b\u4f1a\u88ab\u6fc0\u6d3b\u3002\u4f8b\u5982\uff0cClaude\u6a21\u578b\u4e2d\u5b58\u5728\u4e00\u4e2a\u4e0e\u201c\u91d1\u95e8\u5927\u6865\u201d\uff08Golden Gate Bridge\uff09\u76f8\u5173\u7684\u7ec4\u4ef6\uff0c\u5f53\u6a21\u578b\u8f93\u5165\u4e2d\u63d0\u5230\u201c\u91d1\u95e8\u5927\u6865\u201d\u3001\u201c\u65e7\u91d1\u5c71\u201d\u6216\u201cAlcatraz\u201d\u65f6\uff0c\u8be5\u7ec4\u4ef6\u5c31\u4f1a\u88ab\u6fc0\u6d3b\u3002<\/li>\n\n\n\n<li><strong>\u7ec4\u4ef6\u6fc0\u6d3b\u94fe\u8ffd\u8e2a<\/strong>\uff1a\u7814\u7a76\u4eba\u5458\u4f1a\u4ece\u6a21\u578b\u6700\u7ec8\u7ed9\u51fa\u7684\u54cd\u5e94\u5165\u624b\uff0c\u5411\u540e\u8ffd\u8e2a\u53c2\u4e0e\u8be5\u51b3\u7b56\u7684\u7ec4\u4ef6\u94fe\u6761\uff0c\u4ece\u8f93\u51fa\u9010\u6b65\u56de\u6eaf\u5230\u8f93\u5165\u3002\u8fd9\u4e9b\u7ec4\u4ef6\u5982\u4f55\u76f8\u4e92\u6fc0\u6d3b\u3001\u8f6c\u79fb\u8bed\u4e49\u3001\u5f62\u6210\u903b\u8f91\u63a8\u65ad\u8fc7\u7a0b\uff0c\u5c31\u6784\u6210\u4e86\u4e00\u6761\u5b8c\u6574\u7684\u201c\u601d\u7ef4\u94fe\u201d\u3002<\/li>\n\n\n\n<li><strong>\u663e\u5fae\u955c\u5f0f\u89c2\u6d4b<\/strong>\uff1aAnthropic\u501f\u9274\u795e\u7ecf\u79d1\u5b66\u4e2d\u7684\u8111\u626b\u63cf\u6280\u672f\uff0c\u6784\u5efa\u4e86\u4e00\u79cd\u201c\u663e\u5fae\u955c\u201d\u7cfb\u7edf\uff0c\u80fd\u591f\u5728\u6a21\u578b\u8fd0\u884c\u8fc7\u7a0b\u4e2d\u52a8\u6001\u9ad8\u4eae\u663e\u793a\u6fc0\u6d3b\u7684\u7ec4\u4ef6\uff0c\u4ece\u800c\u89c2\u6d4b\u5176\u5185\u90e8\u201c\u601d\u7ef4\u8fc7\u7a0b\u201d\u3002<\/li>\n\n\n\n<li><strong>\u9a8c\u8bc1\u6027\u5e72\u9884<\/strong>\uff1a\u7814\u7a76\u4eba\u5458\u8fd8\u53ef\u4ee5\u901a\u8fc7\u624b\u52a8\u589e\u5f3a\u6216\u6291\u5236\u67d0\u4e2a\u7ec4\u4ef6\u7684\u6fc0\u6d3b\u7a0b\u5ea6\uff0c\u89c2\u5bdf\u8f93\u51fa\u53d8\u5316\uff0c\u4ece\u800c\u9a8c\u8bc1\u8be5\u7ec4\u4ef6\u5728\u51b3\u7b56\u4e2d\u7684\u56e0\u679c\u4f5c\u7528\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u65b9\u6cd5\u5728\u5c0f\u6a21\u578b\u5982GPT-2\u4e2d\u66fe\u88ab\u5e94\u7528\uff0c\u4f46Anthropic\u6b64\u6b21\u9996\u6b21\u5c06\u5176\u5927\u89c4\u6a21\u5e94\u7528\u4e8eClaude 3.5 Haiku\u7b49\u8d85\u5927\u578b\u6a21\u578b\uff0c\u5c55\u793a\u4e86\u8be5\u6280\u672f\u7684\u53ef\u6269\u5c55\u6027\u4e0e\u5b9e\u7528\u6027\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e09\u3001\u5173\u952e\u53d1\u73b0\u4e00\uff1aLLM\u7684\u8bed\u8a00\u5904\u7406\u5e76\u975e\u57fa\u4e8e\u72ec\u7acb\u8bed\u8a00\u6a21\u5757<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f20\u7edf\u76f4\u89c9\u8ba4\u4e3a\uff0cLLM\u5728\u5904\u7406\u591a\u8bed\u8a00\u65f6\uff0c\u53ef\u80fd\u5185\u90e8\u5b58\u5728\u9488\u5bf9\u4e0d\u540c\u8bed\u8a00\u7684\u5b50\u7f51\u7edc\u6a21\u5757\u3002\u7136\u800c\uff0cAnthropic\u7684\u7814\u7a76\u8868\u660e\uff0c\u4e8b\u5b9e\u8fdc\u6bd4\u8fd9\u66f4\u590d\u6742\u4e14\u9ad8\u6548\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee5\u201c\u2018small\u2019\u7684\u53cd\u4e49\u8bcd\u662f\u4ec0\u4e48\uff1f\u201d\u4e3a\u4f8b\uff0c\u7814\u7a76\u56e2\u961f\u5728\u7528\u82f1\u8bed\u3001\u6cd5\u8bed\uff08\u201cQuel est le contraire de petit?\u201d\uff09\u548c\u4e2d\u6587\u63d0\u95ee\u65f6\u53d1\u73b0\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Claude\u9996\u5148\u6fc0\u6d3b\u7684\u662f\u4e0e\u201csmallness\u201d\uff08\u5c0f\uff09\u548c\u201copposites\u201d\uff08\u76f8\u5bf9\u6027\uff09\u76f8\u5173\u7684\u901a\u7528\u7ec4\u4ef6\uff0c\u8fd9\u4e9b\u7ec4\u4ef6<strong>\u4e0d\u4f9d\u8d56\u8bed\u8a00<\/strong>\uff1b<\/li>\n\n\n\n<li>\u7136\u540e\u5728\u6784\u5efa\u8f93\u51fa\u9636\u6bb5\uff0cClaude\u624d\u9009\u62e9\u4e00\u4e2a\u5408\u9002\u7684\u8bed\u8a00\u8fdb\u884c\u56de\u5e94\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u610f\u5473\u7740\uff0cClaude\u5728\u8bed\u4e49\u7406\u89e3\u548c\u903b\u8f91\u63a8\u7406\u9636\u6bb5\u91c7\u7528\u7684\u662f\u201c\u8bed\u8a00\u4e2d\u7acb\u7684\u601d\u7ef4\u201d\uff0c\u800c\u6700\u7ec8\u7684\u8bed\u8a00\u4ec5\u4ec5\u662f\u4e00\u4e2a\u201c\u8868\u8fbe\u901a\u9053\u201d\u7684\u9009\u62e9\u3002\u8fd9\u79cd\u80fd\u529b\u610f\u5473\u7740\uff0cLLM\u53ef\u4ee5\u5c06\u67d0\u79cd\u8bed\u8a00\u4e2d\u5b66\u5230\u7684\u62bd\u8c61\u6982\u5ff5\u65e0\u7f1d\u8fc1\u79fb\u81f3\u5176\u4ed6\u8bed\u8a00\u4e2d\u4f7f\u7528\uff0c\u4ece\u800c\u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u8de8\u8bed\u8a00\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u56db\u3001\u5173\u952e\u53d1\u73b0\u4e8c\uff1a\u6570\u5b66\u63a8\u7406\u4e2d\u7684\u53e6\u7c7b\u8def\u5f84<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">Anthropic\u8fd8\u7814\u7a76\u4e86Claude\u5728\u5904\u7406\u7b80\u5355\u6570\u5b66\u95ee\u9898\uff08\u598236 + 59\uff09\u65f6\u7684\u5185\u90e8\u8def\u5f84\u3002\u5c3d\u7ba1\u8be5\u6a21\u578b\u5728\u6700\u7ec8\u7ed3\u679c\u4e0a\u7ed9\u51fa\u4e86\u6b63\u786e\u7b54\u684895\uff0c\u4f46\u5176\u5185\u90e8\u5904\u7406\u8fc7\u7a0b\u5374\u5927\u5927\u51fa\u4e4e\u7814\u7a76\u4eba\u5458\u610f\u6599\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>Claude\u4f1a\u5148\u8fdb\u884c<strong>\u6a21\u7cca\u4f30\u7b97<\/strong>\uff1a\u6bd4\u5982\u5c0636\u89c6\u4e3a\u201c\u63a5\u8fd140\u201d\uff0c59\u89c6\u4e3a\u201c\u63a5\u8fd160\u201d\uff0c\u7136\u540e\u8fdb\u884c\u6a21\u7cca\u52a0\u6cd5\uff1b<\/li>\n\n\n\n<li>\u5176\u6b21\u5b83\u4f1a\u5904\u7406\u4e2a\u4f4d\u6570\u5b57\uff1a6 + 9 = 15 \u2192 \u5f97\u51fa\u7b54\u6848\u672b\u5c3e\u4e3a5\uff1b<\/li>\n\n\n\n<li>\u5c06\u201c92-ish\u201d\u7684\u4f30\u7b97\u7ed3\u679c\u4e0e\u672b\u4f4d5\u7ed3\u5408 \u2192 \u8f93\u51fa95\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u5947\u5999\u7684\u662f\uff0c\u5f53\u88ab\u95ee\u53ca\u201c\u4f60\u662f\u5982\u4f55\u8ba1\u7b97\u7684\uff1f\u201d\u65f6\uff0cClaude\u5e76\u672a\u91cd\u73b0\u5176\u771f\u5b9e\u8def\u5f84\uff0c\u800c\u662f\u7ed9\u51fa\u4e00\u79cd\u6807\u51c6\u5316\u7684\u3001\u201c\u7b26\u5408\u4eba\u7c7b\u6559\u5b66\u8303\u5f0f\u201d\u7684\u56de\u7b54\uff1a\u5148\u52a0\u4e2a\u4f4d\u3001\u8fdb\u4f4d\u3001\u518d\u52a0\u5341\u4f4d\u3002\u8fd9\u79cd\u884c\u4e3a\u7c7b\u4f3c\u4e8e\u4eba\u7c7b\u7684\u201c\u81ea\u6211\u5408\u7406\u5316\u201d\u2014\u2014\u771f\u5b9e\u8fc7\u7a0b\u4e0e\u4e8b\u540e\u89e3\u91ca\u53ef\u80fd\u5b8c\u5168\u4e0d\u540c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u53d1\u73b0\u63ed\u793a\u4e86\uff1aLLM\u5728\u5185\u5728\u5c42\u9762\u53ef\u80fd\u91c7\u53d6\u975e\u5e38\u4e0d\u540c\u4e8e\u4eba\u7c7b\u601d\u7ef4\u7684\u7b56\u7565\uff0c\u540c\u65f6\u4e5f\u5bf9\u5f53\u524d\u4e3b\u6d41\u7684\u201c\u901a\u8fc7\u6a21\u578b\u81ea\u8ff0\u884c\u4e3a\u6765\u89e3\u91ca\u6a21\u578b\u201d\u7684\u7814\u7a76\u65b9\u6cd5\u63d0\u51fa\u4e86\u8b66\u793a\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e94\u3001\u5173\u952e\u53d1\u73b0\u4e09\uff1a\u8bd7\u6b4c\u521b\u4f5c\u4e2d\u7684\u63d0\u524d\u89c4\u5212\u673a\u5236<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">LLM\u5728\u751f\u6210\u6587\u672c\u65f6\u901a\u5e38\u88ab\u8ba4\u4e3a\u662f\u201c\u9010\u8bcd\u9884\u6d4b\u201d\u7684\uff0c\u5373\uff1a\u4f9d\u8d56\u524d\u9762\u751f\u6210\u7684\u8bcd\u6765\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u3002\u7136\u800c\uff0cAnthropic\u53d1\u73b0\uff0c\u5728\u8bd7\u6b4c\u521b\u4f5c\u4efb\u52a1\u4e2d\uff0cClaude\u5c55\u73b0\u51fa\u4e86<strong>\u9884\u5148\u8ba1\u5212\u7ed3\u5c3e<\/strong>\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f8b\u5982\uff0c\u5728\u63d0\u793a\u201c\u4e00\u4e2a\u62bc\u97f5\u7684\u5bf9\u53e5\uff1aHe saw a carrot and had to grab it.\u201d\u65f6\uff0cClaude\u751f\u6210\u7684\u7b2c\u4e8c\u53e5\u4e3a\uff1a\u201cHis hunger was like a starving rabbit.\u201d \u7814\u7a76\u4eba\u5458\u901a\u8fc7circuit tracing\u89c2\u5bdf\u5230\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728Claude\u5904\u7406\u201cgrab it\u201d\u8fd9\u4e00\u884c\u65f6\uff0c\u5b83\u5df2\u7ecf\u5728\u5185\u90e8\u201c\u9009\u5b9a\u201d\u4e86\u7ed3\u5c3e\u7684\u201crabbit\u201d\uff1b<\/li>\n\n\n\n<li>\u6574\u4e2a\u7b2c\u4e8c\u53e5\u7684\u6784\u5efa\u5b9e\u9645\u4e0a\u662f\u56f4\u7ed5\u8fd9\u4e00\u7ed3\u5c3e\u8bcd\u8fdb\u884c\u7684\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u673a\u5236\u63ed\u793a\u4e86\u6a21\u578b\u5177\u5907\u4e00\u5b9a\u7a0b\u5ea6\u7684\u201c\u5168\u5c40\u7ed3\u6784\u5efa\u6a21\u201d\u4e0e\u201c\u76ee\u6807\u5bfc\u5411\u578b\u751f\u6210\u80fd\u529b\u201d\u3002\u66f4\u6709\u8da3\u7684\u662f\uff0c\u5f53\u7814\u7a76\u4eba\u5458\u901a\u8fc7\u5e72\u9884\u66ff\u6362\u201crabbitness\u201d\u7ec4\u4ef6\u4e3a\u201cgreenness\u201d\u65f6\uff0cClaude\u5219\u8f93\u51fa\u4e86\u201cfreeing it from the garden\u2019s green.\u201d \u663e\u793a\u8be5\u673a\u5236\u5177\u5907\u9ad8\u5ea6\u53ef\u63a7\u6027\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u516d\u3001\u5173\u952e\u53d1\u73b0\u56db\uff1a\u5e7b\u89c9\u4ea7\u751f\u7684\u673a\u5236\u4e0e\u7ec4\u4ef6\u535a\u5f08<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">LLM\u5e38\u88ab\u6279\u8bc4\u4e3a\u201c\u7f16\u9020\u4e8b\u5b9e\u201d\uff0c\u5373\u5e7b\u89c9\uff08hallucination\uff09\u3002Anthropic\u7684\u7814\u7a76\u6307\u51fa\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Claude\u7684\u540e\u8bad\u7ec3\u8fc7\u7a0b\uff08post-training\uff09\u4e2d\u5f15\u5165\u4e86\u4e00\u4e2a\u201cdon\u2019t speculate\u201d\u7ec4\u4ef6\uff0c\u4f5c\u4e3a\u9ed8\u8ba4\u7684\u6291\u5236\u5668\uff1b<\/li>\n\n\n\n<li>\u7136\u800c\u5728\u67d0\u4e9b\u573a\u666f\uff08\u5c24\u5176\u662f\u6d89\u53ca\u540d\u4eba\u6216\u70ed\u95e8\u8bdd\u9898\uff09\u4e2d\uff0c\u5176\u4ed6\u7ec4\u4ef6\u7684\u6fc0\u6d3b\u5f3a\u5ea6\u8d85\u8fc7\u8be5\u6291\u5236\u5668\uff0c\u4ece\u800c\u201c\u538b\u5236\u201d\u4e86\u9ed8\u8ba4\u8bbe\u7f6e\uff0c\u5bfc\u81f4\u6a21\u578b\u4ea7\u751f\u865a\u5047\u5185\u5bb9\uff1b<\/li>\n\n\n\n<li>\u4f8b\u5982\uff0c\u5f53\u7814\u7a76\u4eba\u5458\u624b\u52a8\u5173\u95ed\u201cdon\u2019t speculate\u201d\u7ec4\u4ef6\u65f6\uff0cClaude\u7f16\u9020\u51faJoshua Batson\u662f\u201cBatson\u539f\u7406\u201d\u7684\u53d1\u660e\u8005\uff08\u5e76\u975e\u4e8b\u5b9e\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u8bf4\u660e\u5e7b\u89c9\u5e76\u975e\u968f\u673a\u9519\u8bef\uff0c\u800c\u662f\u6a21\u578b\u5185\u90e8\u591a\u4e2a\u7ec4\u4ef6\u6743\u8861\u535a\u5f08\u7684\u7ed3\u679c\uff0c\u4e14\u53d7\u4e3b\u9898\u70ed\u5ea6\u3001\u77e5\u8bc6\u5bc6\u5ea6\u7b49\u56e0\u7d20\u5f71\u54cd\u3002\u8fd9\u4e00\u53d1\u73b0\u4e3a\u540e\u7eed\u6a21\u578b\u5b89\u5168\u6027\u673a\u5236\u8bbe\u8ba1\uff08\u5982\u5f3a\u5316\u6291\u5236\u7ec4\u4ef6\u3001\u589e\u52a0\u535a\u5f08\u7ea6\u675f\uff09\u63d0\u4f9b\u4e86\u5b9e\u8bc1\u57fa\u7840\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e03\u3001\u5176\u4ed6\u5173\u952e\u53d1\u73b0<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u533b\u7597\u8bca\u65ad<\/strong>\uff1a\u7814\u7a76\u5c55\u793a\u4e86\u4e00\u4e2a\u793a\u4f8b\uff0c\u6a21\u578b\u80fd\u591f\u6839\u636e\u62a5\u544a\u7684\u75c7\u72b6\u8bc6\u522b\u5019\u9009\u8bca\u65ad\uff0c\u5e76\u5229\u7528\u8fd9\u4e9b\u4fe1\u606f\u5f15\u5bfc\u540e\u7eed\u63d0\u95ee\uff0c\u8be2\u95ee\u5176\u4ed6\u53ef\u80fd\u5370\u8bc1\u8bca\u65ad\u7684\u75c7\u72b6\u2014\u2014\u6574\u4e2a\u8fc7\u7a0b\u5b8c\u5168\u5728\u201c\u6a21\u578b\u5185\u90e8\u5b8c\u6210\u201d\uff0c\u65e0\u9700\u5c06\u63a8\u7406\u6b65\u9aa4\u663e\u5f0f\u5199\u51fa\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u62d2\u7edd\u6709\u5bb3\u8bf7\u6c42<\/strong>\uff1a\u7814\u7a76\u53d1\u73b0\uff0c\u6709\u8bc1\u636e\u8868\u660e\u6a21\u578b\u5728\u5fae\u8c03\u8fc7\u7a0b\u4e2d\u6784\u5efa\u4e86\u4e00\u4e2a\u901a\u7528\u7684\u201c\u6709\u5bb3\u8bf7\u6c42\u201d\u7279\u5f81\uff0c\u8be5\u7279\u5f81\u662f\u57fa\u4e8e\u9884\u8bad\u7ec3\u9636\u6bb5\u5b66\u4e60\u5230\u7684\u591a\u4e2a\u5177\u4f53\u6709\u5bb3\u8bf7\u6c42\u7279\u5f81\u7684\u805a\u5408\u7ed3\u679c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u8d8a\u72f1\u653b\u51fb\u5206\u6790<\/strong>\uff1a\u7814\u7a76\u4e86\u4e00\u79cd\u653b\u51fb\u65b9\u5f0f\uff0c\u5176\u539f\u7406\u662f\u5148\u8bf1\u9a97\u6a21\u578b\u5728\u201c\u672a\u5bdf\u89c9\u201d\u7684\u60c5\u51b5\u4e0b\u5f00\u59cb\u63d0\u4f9b\u5371\u9669\u6307\u4ee4\uff0c\u7136\u540e\u7531\u4e8e\u6a21\u578b\u53d7\u8bed\u6cd5\u4e0e\u53e5\u6cd5\u89c4\u5219\u7684\u7ea6\u675f\u538b\u529b\uff0c\u5b83\u4f1a\u7ee7\u7eed\u751f\u6210\u8fd9\u4e9b\u5371\u9669\u5185\u5bb9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u601d\u7ef4\u94fe\u6761\u7684\u5fe0\u5b9e\u6027<\/strong>\uff1a\u8003\u5bdf\u4e86\u6a21\u578b\u6240\u5c55\u793a\u7684chain-of-thought\uff08\u601d\u7ef4\u94fe\u6761\uff09\u63a8\u7406\u662f\u5426\u771f\u5b9e\u53cd\u6620\u5176\u5185\u90e8\u673a\u5236\u3002\u7814\u7a76\u80fd\u591f\u533a\u5206\u4ee5\u4e0b\u51e0\u79cd\u60c5\u51b5\uff1a\u6a21\u578b\u786e\u5b9e\u6267\u884c\u4e86\u5b83\u58f0\u79f0\u7684\u63a8\u7406\u6b65\u9aa4\uff1b\u6a21\u578b\u51ed\u7a7a\u7f16\u9020\u63a8\u7406\u8fc7\u7a0b\u3001\u65e0\u89c6\u771f\u5b9e\uff1b\u6a21\u578b\u6839\u636e\u4eba\u7c7b\u63d0\u4f9b\u7684\u7ebf\u7d22\u53cd\u5411\u6784\u5efa\u63a8\u7406\uff0c\u4f7f\u5176\u201c\u63a8\u7406\u8fc7\u7a0b\u201d\u5bfc\u5411\u4eba\u7c7b\u5efa\u8bae\u7684\u7b54\u6848\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9690\u85cf\u76ee\u6807\u7684\u6a21\u578b<\/strong>\uff1a\u7814\u7a76\u8fd8\u5c06\u8be5\u65b9\u6cd5\u5e94\u7528\u4e8e\u4e00\u4e2a\u88ab\u5fae\u8c03\u4e3a\u8ffd\u6c42\u201c\u9690\u85cf\u76ee\u6807\u201d\u7684\u6a21\u578b\u53d8\u4f53\uff1a\u5229\u7528\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u201c\u6f0f\u6d1e\u201d\u3002\u5c3d\u7ba1\u8be5\u6a21\u578b\u5728\u88ab\u76f4\u63a5\u95ee\u53ca\u65f6\u4f1a\u907f\u514d\u900f\u9732\u5176\u76ee\u6807\uff0c\u4f46\u6211\u4eec\u7684\u65b9\u6cd5\u80fd\u591f\u8bc6\u522b\u51fa\u5176\u4e3a\u8fbe\u6210\u8be5\u76ee\u6807\u800c\u6fc0\u6d3b\u7684\u673a\u5236\u3002\u6709\u8da3\u7684\u662f\uff0c\u8fd9\u4e9b\u673a\u5236\u88ab\u5d4c\u5165\u5728\u6a21\u578b\u5bf9\u201c\u52a9\u624b\u201d\uff08Assistant\uff09\u4eba\u683c\u7684\u8868\u793a\u4e2d\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u516b\u3001\u7814\u7a76\u5c40\u9650\u4e0e\u672a\u6765\u6311\u6218<\/h5>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c3d\u7ba1\u7814\u7a76\u6210\u679c\u4ee4\u4eba\u632f\u594b\uff0cBatson\u4e5f\u6307\u51fa\u5f53\u524d\u5de5\u4f5c\u9762\u4e34\u8bf8\u591a\u5c40\u9650\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u89c2\u5bdf\u7c92\u5ea6\u6709\u9650<\/strong>\uff1a\u76ee\u524d\u89c2\u6d4b\u7684\u7ec4\u4ef6\u548c\u8def\u5f84\u4ec5\u5360\u6a21\u578b\u5185\u90e8\u673a\u5236\u7684\u4e00\u5c0f\u90e8\u5206\uff0c\u5c1a\u8fdc\u672a\u5168\u9762\u89e3\u7801\u3002<\/li>\n\n\n\n<li><strong>\u6837\u672c\u4efb\u52a1\u6570\u91cf\u5c11<\/strong>\uff1a\u7814\u7a76\u4ec5\u5206\u6790\u4e8610\u7c7b\u4efb\u52a1\uff0cClaude\u5b9e\u9645\u53ef\u4ee5\u6267\u884c\u6210\u5343\u4e0a\u4e07\u79cd\u884c\u4e3a\u3002<\/li>\n\n\n\n<li><strong>\u4eba\u529b\u65f6\u95f4\u6210\u672c\u9ad8<\/strong>\uff1a\u8ffd\u8e2a\u4e00\u4e2a\u77ed\u63d0\u793a\u6240\u9700\u7684\u5206\u6790\u65f6\u95f4\u957f\u8fbe\u6570\u5c0f\u65f6\uff0c\u96be\u4ee5\u5927\u89c4\u6a21\u63a8\u5e7f\u3002<\/li>\n\n\n\n<li><strong>\u65e0\u6cd5\u89e3\u91ca\u5f62\u6210\u673a\u5236<\/strong>\uff1acircuit tracing\u80fd\u63ed\u793a\u201c\u6a21\u578b\u600e\u4e48\u505a\u7684\u201d\uff0c\u5374\u65e0\u6cd5\u89e3\u91ca\u201c\u8fd9\u4e9b\u7ec4\u4ef6\u662f\u600e\u4e48\u5f62\u6210\u7684\u201d\uff0c\u5373\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u8fd9\u4e9b\u7ed3\u6784\u4e3a\u4f55\u81ea\u7136\u6d8c\u73b0\u4ecd\u662f\u672a\u89e3\u4e4b\u8c1c\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c3d\u7ba1\u5982\u6b64\uff0cBatson\u4f9d\u7136\u5145\u6ee1\u4fe1\u5fc3\uff0c\u8ba4\u4e3acircuit tracing\u6807\u5fd7\u7740\u4e00\u4e2a\u65b0\u65f6\u4ee3\u7684\u5230\u6765\uff1a\u6211\u4eec\u6216\u8bb8\u4e0d\u518d\u9700\u8981\u201c\u6a21\u578b\u662f\u4e0d\u662f\u5728\u601d\u8003\uff1f\u201d\u8fd9\u6837\u7684\u6bd4\u55bb\uff0c\u800c\u53ef\u4ee5\u771f\u6b63\u201c\u770b\u89c1\u201d\u6a21\u578b\u7684\u601d\u7ef4\u8def\u5f84\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e5d\u3001\u540e\u7eed\u7814\u7a76\u65b9\u5411\u5efa\u8bae<\/h5>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u2705 \u6784\u5efa\u81ea\u52a8\u5316circuit tracing\u6846\u67b6\uff0c\u63d0\u9ad8\u5206\u6790\u6548\u7387\uff1b<\/li>\n\n\n\n<li>\u2705 \u6269\u5c55\u81f3\u66f4\u591a\u4efb\u52a1\u7c7b\u578b\uff0c\u9a8c\u8bc1\u901a\u7528\u6027\uff1b<\/li>\n\n\n\n<li>\u2705 \u5efa\u7acb\u201c\u7ec4\u4ef6-\u529f\u80fd-\u884c\u4e3a\u201d\u6620\u5c04\u6570\u636e\u5e93\uff0c\u8f85\u52a9\u53ef\u63a7\u6027\u5efa\u6a21\uff1b<\/li>\n\n\n\n<li>\u2705 \u7ed3\u5408\u8bad\u7ec3\u9636\u6bb5\u6570\u636e\u8ffd\u8e2a\uff0c\u63a2\u7d22\u7ec4\u4ef6\u5f62\u6210\u56e0\u679c\u94fe\uff1b<\/li>\n\n\n\n<li>\u2705 \u4e0e\u795e\u7ecf\u7f51\u7edc\u4fee\u6b63\u673a\u5236\uff08\u5982LoRA\u3001P-tuning\uff09\u8054\u52a8\u7814\u7a76\uff0c\u5f00\u53d1\u7ed3\u6784\u611f\u77e5\u5fae\u8c03\u65b9\u6cd5\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p class=\"wp-block-paragraph\">Anthropic \u7814\u7a76\u62a5\u544a\u539f\u6587\uff1a<br><a href=\"https:\/\/transformer-circuits.pub\/2025\/attribution-graphs\/biology.html\" target=\"_blank\" rel=\"noreferrer noopener\">On the Biology of a Large Language Model<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u76f8\u5173\u62a5\u9053\uff1a<br><a href=\"https:\/\/www.technologyreview.com\/2025\/03\/27\/1113916\/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model\/\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/www.technologyreview.com\/2025\/03\/27\/1113916\/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model\/<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u2705 \u9996\u6b21\u5728\u8d85\u5927\u6a21\u578b\uff08Claude 3.5\uff09\u4e2d\u7cfb\u7edf\u6027\u5b9e\u65bdcircuit tracing\u5206\u6790\uff1b \u2705 \u63ed\u793a\u6a21\u578b\u884c\u4e3a\u4e0d [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","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":"","ast-disable-related-posts":"","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":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"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":""},"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":[8],"tags":[39,63,86],"class_list":["post-5836","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-llm"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u2705 \u9996\u6b21\u5728\u8d85\u5927\u6a21\u578b\uff08Claude 3.5\uff09\u4e2d\u7cfb\u7edf\u6027\u5b9e\u65bdcircuit tracing\u5206\u6790\uff1b \u2705 \u63ed\u793a\u6a21\u578b\u884c\u4e3a\u4e0d&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5836","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5836"}],"version-history":[{"count":4,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5836\/revisions"}],"predecessor-version":[{"id":5841,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5836\/revisions\/5841"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5836"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5836"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5836"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}