{"id":34919,"date":"2024-12-15T10:59:10","date_gmt":"2024-12-15T02:59:10","guid":{"rendered":"https:\/\/17aitech.com\/?p=34919"},"modified":"2024-12-15T10:59:10","modified_gmt":"2024-12-15T02:59:10","slug":"llm%e5%8f%af%e8%a7%a3%e9%87%8a%e6%80%a7%e7%9a%84%e6%9c%aa%e6%9d%a5%e5%b8%8c%e6%9c%9b%ef%bc%9f%e7%a8%80%e7%96%8f%e8%87%aa%e7%bc%96%e7%a0%81%e5%99%a8%e6%98%af%e5%a6%82%e4%bd%95%e5%b7%a5%e4%bd%9c","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34919","title":{"rendered":"LLM\u53ef\u89e3\u91ca\u6027\u7684\u672a\u6765\u5e0c\u671b\uff1f\u7a00\u758f\u81ea\u7f16\u7801\u5668\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8fd9\u91cc\u6709\u4e00\u4efd\u76f4\u89c2\u8bf4\u660e"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-08-05-5\" target=\"_blank\">LLM\u53ef\u89e3\u91ca\u6027\u7684\u672a\u6765\u5e0c\u671b\uff1f\u7a00\u758f\u81ea\u7f16\u7801\u5668\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8fd9\u91cc\u6709\u4e00\u4efd\u76f4\u89c2\u8bf4\u660e<\/a><\/p>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"17\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\n<p>\u7b80\u800c\u8a00\u4e4b\uff1a\u77e9\u9635 \u2192 ReLU \u6fc0\u6d3b \u2192 \u77e9\u9635<\/p>\n<\/section>\n<\/section>\n<\/blockquote>\n<p>\u5728\u89e3\u91ca<mark data-type=\"tech_methods\" data-id=\"1a0e9c5e-6502-4cd7-8683-6b5ca6c48be2\">\u673a\u5668\u5b66\u4e60<\/mark>\u6a21\u578b\u65b9\u9762\uff0c<mark data-type=\"tech_methods\" data-id=\"df5b0394-985c-43f8-8063-9d8137501ffd\">\u7a00\u758f\u81ea\u7f16\u7801\u5668<\/mark>\uff08SAE\uff09\u662f\u4e00\u79cd\u8d8a\u6765\u8d8a\u5e38\u7528\u7684\u5de5\u5177\uff08\u867d\u7136 SAE \u5728 1997 \u5e74\u5de6\u53f3\u5c31\u5df2\u7ecf\u95ee\u4e16\u4e86\uff09\u3002<\/p>\n<p><mark data-type=\"tech_methods\" data-id=\"1a0e9c5e-6502-4cd7-8683-6b5ca6c48be2\">\u673a\u5668\u5b66\u4e60<\/mark>\u6a21\u578b\u548c LLM \u6b63\u53d8\u5f97\u8d8a\u6765\u8d8a\u5f3a\u5927\u3001\u8d8a\u6765\u8d8a\u6709\u7528\uff0c\u4f46\u5b83\u4eec\u4ecd\u65e7\u662f\u9ed1\u7bb1\uff0c\u6211\u4eec\u5e76\u4e0d\u7406\u89e3\u5b83\u4eec\u5b8c\u6210\u4efb\u52a1\u7684\u65b9\u5f0f\u3002\u7406\u89e3\u5b83\u4eec\u7684\u5de5\u4f5c\u65b9\u5f0f\u5e94\u5f53\u5927\u6709\u52a9\u76ca\u3002<\/p>\n<p>SAE \u53ef\u5e2e\u52a9\u6211\u4eec\u5c06\u6a21\u578b\u7684\u8ba1\u7b97\u5206\u89e3\u6210\u53ef\u4ee5\u7406\u89e3\u7684\u7ec4\u4ef6\u3002\u8fd1\u65e5\uff0cLLM \u53ef\u89e3\u91ca\u6027\u7814\u7a76\u8005 Adam Karvonen \u53d1\u5e03\u4e86\u4e00\u7bc7\u535a\u5ba2\u6587\u7ae0\uff0c\u76f4\u89c2\u5730\u89e3\u91ca\u4e86 SAE \u7684\u5de5\u4f5c\u65b9\u5f0f\u3002<\/p>\n<p><strong>\u53ef\u89e3\u91ca\u6027\u7684\u96be\u9898<\/strong><\/p>\n<p><mark data-type=\"tech_methods\" data-id=\"72b0bcc0-d8f9-4edd-919f-fa7c2560388c\">\u795e\u7ecf\u7f51\u7edc<\/mark>\u6700\u81ea\u7136\u7684\u7ec4\u4ef6\u662f\u5404\u4e2a<mark data-type=\"concepts\" data-id=\"b5cce590-1c61-4d22-8e75-fe74128079c3\">\u795e\u7ecf\u5143<\/mark>\u3002\u4e0d\u5e78\u7684\u662f\uff0c\u5355\u4e2a<mark data-type=\"concepts\" data-id=\"b5cce590-1c61-4d22-8e75-fe74128079c3\">\u795e\u7ecf\u5143<\/mark>\u5e76\u4e0d\u80fd\u4fbf\u6377\u5730\u4e0e\u5355\u4e2a\u6982\u5ff5\u76f8\u5bf9\u5e94\uff0c\u6bd4\u5982\u5b66\u672f\u5f15\u7528\u3001\u82f1\u8bed\u5bf9\u8bdd\u3001HTTP \u8bf7\u6c42\u548c\u97e9\u8bed\u6587\u672c\u3002\u5728<mark data-type=\"tech_methods\" data-id=\"72b0bcc0-d8f9-4edd-919f-fa7c2560388c\">\u795e\u7ecf\u7f51\u7edc<\/mark>\u4e2d\uff0c\u6982\u5ff5\u662f\u901a\u8fc7<mark data-type=\"concepts\" data-id=\"b5cce590-1c61-4d22-8e75-fe74128079c3\">\u795e\u7ecf\u5143<\/mark>\u7684\u7ec4\u5408\u8868\u793a\u7684\uff0c\u8fd9\u88ab\u79f0\u4e3a\u53e0\u52a0\uff08superposition\uff09\u3002<\/p>\n<p>\u4e4b\u6240\u4ee5\u4f1a\u8fd9\u6837\uff0c\u662f\u56e0\u4e3a\u4e16\u754c\u4e0a\u5f88\u591a\u53d8\u91cf\u5929\u7136\u5c31\u662f\u7a00\u758f\u7684\u3002<\/p>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u67d0\u4f4d\u540d\u4eba\u7684\u51fa\u751f\u5730\u53ef\u80fd\u51fa\u73b0\u5728\u4e0d\u5230\u5341\u4ebf\u5206\u4e4b\u4e00\u7684\u8bad\u7ec3 token \u4e2d\uff0c\u4f46\u73b0\u4ee3 LLM \u4f9d\u7136\u80fd\u5b66\u5230\u8fd9\u4e00\u4e8b\u5b9e\u4ee5\u53ca\u6709\u5173\u8fd9\u4e2a\u4e16\u754c\u7684\u5927\u91cf\u5176\u5b83\u77e5\u8bc6\u3002\u8bad\u7ec3\u6570\u636e\u4e2d\u5355\u4e2a\u4e8b\u5b9e\u548c\u6982\u5ff5\u7684\u6570\u91cf\u591a\u4e8e\u6a21\u578b\u4e2d<mark data-type=\"concepts\" data-id=\"b5cce590-1c61-4d22-8e75-fe74128079c3\">\u795e\u7ecf\u5143<\/mark>\u7684\u6570\u91cf\uff0c\u8fd9\u53ef\u80fd\u5c31\u662f\u53e0\u52a0\u51fa\u73b0\u7684\u539f\u56e0\u3002<\/p>\n<p>\u8fd1\u6bb5\u65f6\u95f4\uff0c<mark data-type=\"tech_methods\" data-id=\"df5b0394-985c-43f8-8063-9d8137501ffd\">\u7a00\u758f\u81ea\u7f16\u7801\u5668<\/mark>\uff08SAE\uff09\u6280\u672f\u8d8a\u6765\u8d8a\u5e38\u88ab\u7528\u4e8e\u5c06<mark data-type=\"tech_methods\" data-id=\"72b0bcc0-d8f9-4edd-919f-fa7c2560388c\">\u795e\u7ecf\u7f51\u7edc<\/mark>\u5206\u89e3\u6210\u53ef\u7406\u89e3\u7684\u7ec4\u4ef6\u3002SAE \u7684\u8bbe\u8ba1\u7075\u611f\u6765\u81ea<mark data-type=\"concepts\" data-id=\"306453a1-dae0-4518-9204-c5d533cedbe7\">\u795e\u7ecf\u79d1\u5b66<\/mark>\u9886\u57df\u7684\u7a00\u758f\u7f16\u7801\u5047\u8bbe\u3002\u73b0\u5728\uff0cSAE \u5df2\u6210\u4e3a\u89e3\u8bfb\u4eba\u5de5<mark data-type=\"tech_methods\" data-id=\"72b0bcc0-d8f9-4edd-919f-fa7c2560388c\">\u795e\u7ecf\u7f51\u7edc<\/mark>\u65b9\u9762\u6700\u6709\u6f5c\u529b\u7684\u5de5\u5177\u4e4b\u4e00\u3002SAE \u4e0e\u6807\u51c6\u81ea\u7f16\u7801\u5668\u7c7b\u4f3c\u3002<\/p>\n<p>\u5e38\u89c4\u81ea\u7f16\u7801\u5668\u662f\u4e00\u79cd\u7528\u4e8e\u538b\u7f29\u5e76\u91cd\u5efa\u8f93\u5165\u6570\u636e\u7684<mark data-type=\"tech_methods\" data-id=\"72b0bcc0-d8f9-4edd-919f-fa7c2560388c\">\u795e\u7ecf\u7f51\u7edc<\/mark>\u3002<\/p>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5982\u679c\u8f93\u5165\u662f\u4e00\u4e2a 100 \u7ef4\u7684\u5411\u91cf\uff08\u5305\u542b 100 \u4e2a\u6570\u503c\u7684\u5217\u8868\uff09\uff1b\u81ea\u7f16\u7801\u5668\u9996\u5148\u4f1a\u8ba9\u8be5\u8f93\u5165\u901a\u8fc7\u4e00\u4e2a\u7f16\u7801\u5668\u5c42\uff0c\u8ba9\u5176\u88ab\u538b\u7f29\u6210\u4e00\u4e2a 50 \u7ef4\u7684\u5411\u91cf\uff0c\u7136\u540e\u5c06\u8fd9\u4e2a\u538b\u7f29\u540e\u7684\u7f16\u7801\u8868\u793a\u9988\u9001\u7ed9\u89e3\u7801\u5668\uff0c\u5f97\u5230 100 \u7ef4\u7684\u8f93\u51fa\u5411\u91cf\u3002\u5176\u91cd\u5efa\u8fc7\u7a0b\u901a\u5e38\u5e76\u4e0d\u5b8c\u7f8e\uff0c\u56e0\u4e3a\u538b\u7f29\u8fc7\u7a0b\u4f1a\u8ba9\u91cd\u5efa\u4efb\u52a1\u53d8\u5f97\u975e\u5e38\u56f0\u96be\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-9b60a79aed716aaa25e35f01a3df08ad.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-9b60a79aed716aaa25e35f01a3df08ad.png\"><\/a><\/p>\n<p><em><sup>\u4e00\u4e2a\u6807\u51c6\u81ea\u7f16\u7801\u5668\u7684\u793a\u610f\u56fe\uff0c\u5176\u6709 1&#215;4 \u7684\u8f93\u5165\u5411\u91cf\u30011&#215;2 \u7684\u4e2d\u95f4\u72b6\u6001\u5411\u91cf\u548c 1&#215;4 \u7684\u8f93\u51fa\u5411\u91cf\u3002\u5355\u5143\u683c\u7684\u989c\u8272\u8868\u793a\u6fc0\u6d3b\u503c\u3002\u8f93\u51fa\u662f\u8f93\u5165\u7684\u4e0d\u5b8c\u7f8e\u91cd\u5efa\u7ed3\u679c\u3002<\/sup><\/em><\/p>\n<p><strong>\u89e3\u91ca<mark data-type=\"tech_methods\" data-id=\"df5b0394-985c-43f8-8063-9d8137501ffd\">\u7a00\u758f\u81ea\u7f16\u7801\u5668<\/mark><\/strong><\/p>\n<p><strong><mark data-type=\"tech_methods\" data-id=\"df5b0394-985c-43f8-8063-9d8137501ffd\">\u7a00\u758f\u81ea\u7f16\u7801\u5668<\/mark>\u7684\u5de5\u4f5c\u65b9\u5f0f<\/strong><\/p>\n<p><mark data-type=\"tech_methods\" data-id=\"df5b0394-985c-43f8-8063-9d8137501ffd\">\u7a00\u758f\u81ea\u7f16\u7801\u5668<\/mark>\u4f1a\u5c06\u8f93\u5165\u5411\u91cf\u8f6c\u6362\u6210\u4e2d\u95f4\u5411\u91cf\uff0c\u8be5\u4e2d\u95f4\u5411\u91cf\u7684\u7ef4\u5ea6\u53ef\u80fd\u9ad8\u4e8e\u3001\u7b49\u4e8e\u6216\u4f4e\u4e8e\u8f93\u5165\u7684\u7ef4\u5ea6\u3002\u5728\u7528\u4e8e LLM \u65f6\uff0c\u4e2d\u95f4\u5411\u91cf\u7684\u7ef4\u5ea6\u901a\u5e38\u9ad8\u4e8e\u8f93\u5165\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u5982\u679c\u4e0d\u52a0\u989d\u5916\u7684\u7ea6\u675f\u6761\u4ef6\uff0c\u90a3\u4e48\u8be5\u4efb\u52a1\u5c31\u5f88\u7b80\u5355\uff0cSAE \u53ef\u4ee5\u4f7f\u7528\u5355\u4f4d\u77e9\u9635\u6765\u5b8c\u7f8e\u5730\u91cd\u5efa\u51fa\u8f93\u5165\uff0c\u4e0d\u4f1a\u51fa\u73b0\u4efb\u4f55\u610f\u6599\u4e4b\u5916\u7684\u4e1c\u897f\u3002\u4f46\u6211\u4eec\u4f1a\u6dfb\u52a0\u7ea6\u675f\u6761\u4ef6\uff0c\u5176\u4e2d\u4e4b\u4e00\u662f\u4e3a\u8bad\u7ec3\u635f\u5931\u6dfb\u52a0\u7a00\u758f\u5ea6\u60e9\u7f5a\uff0c\u8fd9\u4f1a\u4fc3\u4f7f SAE \u521b\u5efa\u7a00\u758f\u7684\u4e2d\u95f4\u5411\u91cf\u3002<\/p>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u6211\u4eec\u53ef\u4ee5\u5c06 100 \u7ef4\u7684\u8f93\u5165\u6269\u5c55\u6210 200 \u7ef4\u7684\u5df2\u7f16\u7801\u8868\u5f81\u5411\u91cf\uff0c\u5e76\u4e14\u6211\u4eec\u53ef\u4ee5\u8bad\u7ec3 SAE \u4f7f\u5176\u5728\u5df2\u7f16\u7801\u8868\u5f81\u4e2d\u4ec5\u6709\u5927\u7ea6 20 \u4e2a\u975e\u96f6\u5143\u7d20\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-937b3c2e202a803defa1aedb1c44c479.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-937b3c2e202a803defa1aedb1c44c479.png\"><\/a><\/p>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <mark data-type=\"tech_methods\" data-id=\"df5b0394-985c-43f8-8063-9d8137501ffd\">\u7a00\u758f\u81ea\u7f16\u7801\u5668<\/mark>\u793a\u610f\u56fe\u3002\u8bf7\u6ce8\u610f\uff0c\u4e2d\u95f4\u6fc0\u6d3b\u662f\u7a00\u758f\u7684\uff0c\u4ec5\u6709 2 \u4e2a\u975e\u96f6\u503c\u3002<\/sup><br \/><\/em><\/p>\n<p>\u6211\u4eec\u5c06 SAE \u7528\u4e8e<mark data-type=\"tech_methods\" data-id=\"72b0bcc0-d8f9-4edd-919f-fa7c2560388c\">\u795e\u7ecf\u7f51\u7edc<\/mark>\u5185\u7684\u4e2d\u95f4\u6fc0\u6d3b\uff0c\u800c<mark data-type=\"tech_methods\" data-id=\"72b0bcc0-d8f9-4edd-919f-fa7c2560388c\">\u795e\u7ecf\u7f51\u7edc<\/mark>\u53ef\u80fd\u5305\u542b\u8bb8\u591a\u5c42\u3002\u5728\u524d\u5411\u901a\u8fc7\u8fc7\u7a0b\u4e2d\uff0c\u6bcf\u4e00\u5c42\u4e2d\u548c\u6bcf\u4e00\u5c42\u4e4b\u95f4\u90fd\u6709\u4e2d\u95f4\u6fc0\u6d3b\u3002<\/p>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0cGPT-3 \u6709 96 \u5c42\u3002\u5728\u524d\u5411\u901a\u8fc7\u8fc7\u7a0b\u4e2d\uff0c\u8f93\u5165\u4e2d\u7684\u6bcf\u4e2a token \u90fd\u6709\u4e00\u4e2a 12,288 \u7ef4\u5411\u91cf\uff08\u4e00\u4e2a\u5305\u542b 12,288 \u4e2a\u6570\u503c\u7684\u5217\u8868\uff09\u3002\u6b64\u5411\u91cf\u4f1a\u7d2f\u79ef\u6a21\u578b\u5728\u6bcf\u4e00\u5c42\u5904\u7406\u65f6\u7528\u4e8e\u9884\u6d4b\u4e0b\u4e00 token \u7684\u6240\u6709\u4fe1\u606f\uff0c\u4f46\u5b83\u5e76\u4e0d\u900f\u660e\uff0c\u8ba9\u4eba\u96be\u4ee5\u7406\u89e3\u5176\u4e2d\u7a76\u7adf\u5305\u542b\u4ec0\u4e48\u4fe1\u606f\u3002<\/p>\n<p>\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528 SAE \u6765\u7406\u89e3\u8fd9\u79cd\u4e2d\u95f4\u6fc0\u6d3b\u3002SAE \u57fa\u672c\u4e0a\u5c31\u662f\u300c\u77e9\u9635 \u2192 ReLU \u6fc0\u6d3b \u2192 \u77e9\u9635\u300d\u3002<\/p>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5982\u679c GPT-3 SAE \u7684\u6269\u5c55\u56e0\u5b50\u4e3a 4\uff0c\u5176\u8f93\u5165\u6fc0\u6d3b\u6709 12,288 \u7ef4\uff0c\u5219\u5176 SAE \u7f16\u7801\u7684\u8868\u5f81\u6709 49,512 \u7ef4\uff0812,288 x 4\uff09\u3002\u7b2c\u4e00\u4e2a\u77e9\u9635\u662f\u5f62\u72b6\u4e3a (12,288, 49,512) \u7684\u7f16\u7801\u5668\u77e9\u9635\uff0c\u7b2c\u4e8c\u4e2a\u77e9\u9635\u662f\u5f62\u72b6\u4e3a (49,512, 12,288) \u7684\u89e3\u7801\u5668\u77e9\u9635\u3002\u901a\u8fc7\u8ba9 GPT \u7684\u6fc0\u6d3b\u4e0e\u7f16\u7801\u5668\u76f8\u4e58\u5e76\u4f7f\u7528 ReLU\uff0c\u53ef\u4ee5\u5f97\u5230 49,512 \u7ef4\u7684 SAE \u7f16\u7801\u7684\u7a00\u758f\u8868\u5f81\uff0c\u56e0\u4e3a SAE \u7684<mark data-type=\"concepts\" data-id=\"4c38563a-2d9b-439e-bfb4-21d209eeff3e\">\u635f\u5931\u51fd\u6570<\/mark>\u4f1a\u4fc3\u4f7f\u5b9e\u73b0\u7a00\u758f\u6027\u3002<\/p>\n<p>\u901a\u5e38\u6765\u8bf4\uff0c\u6211\u4eec\u7684\u76ee\u6807\u8ba9 SAE \u7684\u8868\u5f81\u4e2d\u975e\u96f6\u503c\u7684\u6570\u91cf\u5c11\u4e8e 100 \u4e2a\u3002\u901a\u8fc7\u5c06 SAE \u7684\u8868\u5f81\u4e0e\u89e3\u7801\u5668\u76f8\u4e58\uff0c\u53ef\u5f97\u5230\u4e00\u4e2a 12,288 \u7ef4\u7684\u91cd\u5efa\u7684\u6a21\u578b\u6fc0\u6d3b\u3002\u8fd9\u4e2a\u91cd\u5efa\u7ed3\u679c\u5e76\u4e0d\u80fd\u4e0e\u539f\u59cb\u7684 GPT \u6fc0\u6d3b\u5b8c\u7f8e\u5339\u914d\uff0c\u56e0\u4e3a\u7a00\u758f\u6027\u7ea6\u675f\u6761\u4ef6\u4f1a\u8ba9\u5b8c\u7f8e\u5339\u914d\u96be\u4ee5\u5b9e\u73b0\u3002<\/p>\n<p>\u4e00\u822c\u6765\u8bf4\uff0c\u4e00\u4e2a SAE \u4ec5\u7528\u4e8e\u6a21\u578b\u4e2d\u7684\u4e00\u4e2a\u4f4d\u7f6e\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u6211\u4eec\u53ef\u4ee5\u5728 26 \u548c 27 \u5c42\u4e4b\u95f4\u7684\u4e2d\u95f4\u6fc0\u6d3b\u4e0a\u8bad\u7ec3\u4e00\u4e2a SAE\u3002\u4e3a\u4e86\u5206\u6790 GPT-3 \u7684\u5168\u90e8 96 \u5c42\u7684\u8f93\u51fa\u4e2d\u5305\u542b\u7684\u4fe1\u606f\uff0c\u53ef\u4ee5\u8bad\u7ec3 96 \u4e2a\u5206\u7acb\u7684 SAE\u2014\u2014 \u6bcf\u5c42\u7684\u8f93\u51fa\u90fd\u6709\u4e00\u4e2a\u3002\u5982\u679c\u6211\u4eec\u4e5f\u60f3\u5206\u6790\u6bcf\u4e00\u5c42\u5185\u5404\u79cd\u4e0d\u540c\u7684\u4e2d\u95f4\u6fc0\u6d3b\uff0c\u90a3\u5c31\u9700\u8981\u6570\u767e\u4e2a SAE\u3002\u4e3a\u4e86\u83b7\u53d6\u8fd9\u4e9b SAE \u7684\u8bad\u7ec3\u6570\u636e\uff0c\u9700\u8981\u5411\u8fd9\u4e2a GPT \u6a21\u578b\u8f93\u5165\u5927\u91cf\u4e0d\u540c\u7684\u6587\u672c\uff0c\u7136\u540e\u6536\u96c6\u6bcf\u4e2a\u9009\u5b9a\u4f4d\u7f6e\u7684\u4e2d\u95f4\u6fc0\u6d3b\u3002<\/p>\n<p>\u4e0b\u9762\u63d0\u4f9b\u4e86\u4e00\u4e2a SAE \u7684 PyTorch \u53c2\u8003\u5b9e\u73b0\u3002\u5176\u4e2d\u7684\u53d8\u91cf\u5e26\u6709\u5f62\u72b6\u6ce8\u91ca\uff0c\u8fd9\u4e2a\u70b9\u5b50\u6765\u81ea Noam Shazeer\uff0c\u53c2\u89c1\uff1ahttps:\/\/medium.com\/@NoamShazeer\/shape-suffixes-good-coding-style-f836e72e24fd \u3002\u8bf7\u6ce8\u610f\uff0c\u4e3a\u4e86\u5c3d\u53ef\u80fd\u5730\u63d0\u5347\u6027\u80fd\uff0c\u4e0d\u540c\u7684 SAE \u5b9e\u73b0\u5f80\u5f80\u4f1a\u6709\u4e0d\u540c\u7684\u504f\u7f6e\u9879\u3001\u5f52\u4e00\u5316\u65b9\u6848\u6216\u521d\u59cb\u5316\u65b9\u6848\u3002\u6700\u5e38\u89c1\u7684\u4e00\u79cd\u9644\u52a0\u9879\u662f\u67d0\u79cd\u5bf9\u89e3\u7801\u5668\u5411\u91cf<mark data-type=\"concepts\" data-id=\"0caf290b-4ff5-445b-8f62-ecc75a6a237f\">\u8303\u6570<\/mark>\u7684\u7ea6\u675f\u3002\u66f4\u591a\u7ec6\u8282\u8bf7\u8bbf\u95ee\u4ee5\u4e0b\u5b9e\u73b0\uff1a<\/p>\n<ul>\n<li>\n<p>OpenAI\uff1ahttps:\/\/github.com\/openai\/sparse_autoencoder\/blob\/main\/sparse_autoencoder\/model.py#L16<\/p>\n<\/li>\n<li>\n<p>SAELens\uff1ahttps:\/\/github.com\/jbloomAus\/SAELens\/blob\/main\/sae_lens\/sae.py#L97<\/p>\n<\/li>\n<li>\n<p>dictionary_learning\uff1ahttps:\/\/github.com\/saprmarks\/dictionary_learning\/blob\/main\/dictionary.py#L30<\/p>\n<\/li>\n<\/ul>\n<section>\n<pre data-lang=\"python\"><code>import torch<\/code>\r\n<code>import\u00a0torch.nn\u00a0as\u00a0nn<\/code>\r\n<code>\r\n<\/code><code># D = d_model, F = dictionary_size<\/code>\r\n<code># e.g. if d_model = 12288 and dictionary_size = 49152<\/code>\r\n<code>#\u00a0then\u00a0model_activations_D.shape\u00a0=\u00a0(12288,)\u00a0and\u00a0encoder_DF.weight.shape\u00a0=\u00a0(12288,\u00a049152)<\/code>\r\n<code>\r\n<\/code><code>class SparseAutoEncoder (nn.Module):<\/code>\r\n<code>    \"\"\"<\/code>\r\n<code>    A one-layer autoencoder.<\/code>\r\n<code>    \"\"\"<\/code>\r\n<code>    def __init__(self, activation_dim: int, dict_size: int):<\/code>\r\n<code>        super ().__init__()<\/code>\r\n<code>        self.activation_dim = activation_dim<\/code>\r\n<code>        self.dict_size = dict_size<\/code><code>\r\n<\/code><code>\r\n<\/code><code>        self.encoder_DF = nn.Linear (activation_dim, dict_size, bias=True)<\/code>\r\n<code>        self.decoder_FD = nn.Linear (dict_size, activation_dim, bias=True)<\/code><code>\r\n<\/code><code>\r\n<\/code><code>    def encode (self, model_activations_D: torch.Tensor) -&gt; torch.Tensor:<\/code>\r\n<code>        return nn.ReLU ()(self.encoder_DF (model_activations_D))<\/code>\r\n<code>\r\n<\/code><code>    def decode (self, encoded_representation_F: torch.Tensor) -&gt; torch.Tensor:<\/code>\r\n<code>        return self.decoder_FD (encoded_representation_F)<\/code>\r\n<code>\r\n<\/code><code>    def forward_pass (self, model_activations_D: torch.Tensor) -&gt; tuple [torch.Tensor, torch.Tensor]:<\/code>\r\n<code>        encoded_representation_F = self.encode (model_activations_D)<\/code>\r\n<code>        reconstructed_model_activations_D = self.decode (encoded_representation_F)<\/code>\r\n<code>\u00a0 \u00a0 \u00a0 \u00a0 return reconstructed_model_activations_D, encoded_representation_F<\/code><\/pre>\n<\/section>\n<p>\u6807\u51c6\u81ea\u7f16\u7801\u5668\u7684<mark data-type=\"concepts\" data-id=\"4c38563a-2d9b-439e-bfb4-21d209eeff3e\">\u635f\u5931\u51fd\u6570<\/mark>\u57fa\u4e8e\u8f93\u5165\u91cd\u5efa\u7ed3\u679c\u7684\u51c6\u786e\u5ea6\u3002\u4e3a\u4e86\u5f15\u5165\u7a00\u758f\u6027\uff0c\u6700\u76f4\u63a5\u7684\u65b9\u6cd5\u662f\u5411 SAE \u7684<mark data-type=\"concepts\" data-id=\"4c38563a-2d9b-439e-bfb4-21d209eeff3e\">\u635f\u5931\u51fd\u6570<\/mark>\u6dfb\u52a0\u4e00\u4e2a\u7a00\u758f\u5ea6\u60e9\u7f5a\u9879\u3002\u5bf9\u4e8e\u8fd9\u4e2a\u60e9\u7f5a\u9879\uff0c\u6700\u5e38\u89c1\u7684\u8ba1\u7b97\u65b9\u5f0f\u662f\u53d6\u8fd9\u4e2a SAE \u7684\u5df2\u7f16\u7801\u8868\u5f81\uff08\u800c\u975e SAE <mark data-type=\"concepts\" data-id=\"149a12cf-10c2-4555-9899-cc6dee319ef5\">\u6743\u91cd<\/mark>\uff09\u7684 L1 \u635f\u5931\u5e76\u5c06\u5176\u4e58\u4ee5\u4e00\u4e2a L1 \u7cfb\u6570\u3002\u8fd9\u4e2a L1 \u7cfb\u6570\u662f SAE \u8bad\u7ec3\u4e2d\u7684\u4e00\u4e2a\u5173\u952e<mark data-type=\"concepts\" data-id=\"5619ca3f-5d4e-48c1-824d-d2a0aea0c7d1\">\u8d85\u53c2\u6570<\/mark>\uff0c\u56e0\u4e3a\u5b83\u53ef\u786e\u5b9a\u5b9e\u73b0\u7a00\u758f\u5ea6\u4e0e\u7ef4\u6301\u91cd\u5efa\u51c6\u786e\u5ea6\u4e4b\u95f4\u7684\u6743\u8861\u3002<\/p>\n<p>\u8bf7\u6ce8\u610f\uff0c\u8fd9\u91cc\u5e76\u6ca1\u6709\u9488\u5bf9\u53ef\u89e3\u91ca\u6027\u8fdb\u884c\u4f18\u5316\u3002\u76f8\u53cd\uff0c\u53ef\u89e3\u91ca\u7684 SAE \u7279\u5f81\u662f\u4f18\u5316\u7a00\u758f\u5ea6\u548c\u91cd\u5efa\u7684\u4e00\u4e2a\u9644\u5e26\u6548\u679c\u3002\u4e0b\u9762\u662f\u4e00\u4e2a\u53c2\u8003<mark data-type=\"concepts\" data-id=\"4c38563a-2d9b-439e-bfb4-21d209eeff3e\">\u635f\u5931\u51fd\u6570<\/mark>\u3002<\/p>\n<section>\n<pre data-lang=\"properties\"><code># B = batch size, D = d_model, F = dictionary_size<\/code>\r\n<code>def calculate_loss (autoencoder: SparseAutoEncoder, model_activations_BD: torch.Tensor, l1_coeffient: float) -&gt; torch.Tensor:<\/code>\r\n<code>    reconstructed_model_activations_BD, encoded_representation_BF = autoencoder.forward_pass (model_activations_BD)<\/code>\r\n<code>    reconstruction_error_BD = (reconstructed_model_activations_BD - model_activations_BD).pow (2)<\/code>\r\n<code>    reconstruction_error_B = einops.reduce (reconstruction_error_BD, 'B D -&gt; B', 'sum')<\/code>\r\n<code>    l2_loss = reconstruction_error_B.mean ()<\/code>\r\n<code>\r\n<\/code><code>    l1_loss = l1_coefficient * encoded_representation_BF.sum ()<\/code>\r\n<code>    loss = l2_loss + l1_loss<\/code>\r\n<code>\u00a0 \u00a0 return loss<\/code><\/pre>\n<\/section>\n<p><em><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-de2349a6aa85dc02cfb52f933056865b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-de2349a6aa85dc02cfb52f933056865b.png\"><\/a><\/em><\/p>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <mark data-type=\"tech_methods\" data-id=\"df5b0394-985c-43f8-8063-9d8137501ffd\">\u7a00\u758f\u81ea\u7f16\u7801\u5668<\/mark>\u7684\u524d\u5411\u901a\u8fc7\u793a\u610f\u56fe\u3002<\/sup><\/em><\/p>\n<p>\u8fd9\u662f<mark data-type=\"tech_methods\" data-id=\"df5b0394-985c-43f8-8063-9d8137501ffd\">\u7a00\u758f\u81ea\u7f16\u7801\u5668<\/mark>\u7684\u5355\u6b21\u524d\u5411\u901a\u8fc7\u8fc7\u7a0b\u3002\u9996\u5148\u662f 1&#215;4 \u5927\u5c0f\u7684\u6a21\u578b\u5411\u91cf\u3002\u7136\u540e\u5c06\u5176\u4e58\u4ee5\u4e00\u4e2a 4&#215;8 \u7684\u7f16\u7801\u5668\u77e9\u9635\uff0c\u5f97\u5230\u4e00\u4e2a 1&#215;8 \u7684\u5df2\u7f16\u7801\u5411\u91cf\uff0c\u7136\u540e\u5e94\u7528 ReLU \u5c06\u8d1f\u503c\u53d8\u6210\u96f6\u3002\u8fd9\u4e2a\u7f16\u7801\u540e\u7684\u5411\u91cf\u5c31\u662f\u7a00\u758f\u7684\u3002\u4e4b\u540e\uff0c\u518d\u8ba9\u5176\u4e58\u4ee5\u4e00\u4e2a 8&#215;4 \u7684\u89e3\u7801\u5668\u77e9\u9635\uff0c\u5f97\u5230\u4e00\u4e2a 1&#215;4 \u7684\u4e0d\u5b8c\u7f8e\u91cd\u5efa\u7684\u6a21\u578b\u6fc0\u6d3b\u3002<\/p>\n<p><strong>\u5047\u60f3\u7684 SAE \u7279\u5f81\u6f14\u793a<\/strong><\/p>\n<p>\u7406\u60f3\u60c5\u51b5\u4e0b\uff0cSAE \u8868\u5f81\u4e2d\u7684\u6bcf\u4e2a\u6709\u6548\u6570\u503c\u90fd\u5bf9\u5e94\u4e8e\u67d0\u4e2a\u53ef\u7406\u89e3\u7684\u7ec4\u4ef6\u3002<\/p>\n<p>\u8fd9\u91cc\u5047\u8bbe\u4e00\u4e2a\u6848\u4f8b\u8fdb\u884c\u8bf4\u660e\u3002\u5047\u8bbe\u4e00\u4e2a 12,288 \u7ef4\u5411\u91cf [1.5, 0.2, -1.2, &#8230;] \u5728 GPT-3 \u770b\u6765\u662f\u8868\u793a\u300cGolden Retriever\u300d\uff08\u91d1\u6bdb\u72ac\uff09\u3002SAE \u662f\u4e00\u4e2a\u5f62\u72b6\u4e3a (49,512, 12,288) \u7684\u77e9\u9635\uff0c\u4f46\u6211\u4eec\u4e5f\u53ef\u4ee5\u5c06\u5176\u770b\u4f5c\u662f 49,512 \u4e2a\u5411\u91cf\u7684\u96c6\u5408\uff0c\u5176\u4e2d\u6bcf\u4e2a\u5411\u91cf\u7684\u5f62\u72b6\u90fd\u662f (1, 12,288)\u3002\u5982\u679c\u8be5 SAE \u89e3\u7801\u5668\u7684 317 \u5411\u91cf\u5b66\u4e60\u5230\u4e86\u4e0e GPT-3 \u90a3\u4e00\u6837\u7684\u300cGolden Retriever\u300d\u6982\u5ff5\uff0c\u90a3\u4e48\u8be5\u89e3\u7801\u5668\u5411\u91cf\u5927\u81f4\u4e5f\u7b49\u4e8e [1.5, 0.2, -1.2, &#8230;]\u3002<\/p>\n<p>\u65e0\u8bba\u4f55\u65f6 SAE \u7684\u6fc0\u6d3b\u7684 317 \u5143\u7d20\u662f\u975e\u96f6\u7684\uff0c\u90a3\u4e48\u5bf9\u5e94\u4e8e\u300cGolden Retriever\u300d\u7684\u5411\u91cf\uff08\u5e76\u6839\u636e 317 \u5143\u7d20\u7684\u5e45\u5ea6\uff09\u4f1a\u88ab\u6dfb\u52a0\u5230\u91cd\u5efa\u6fc0\u6d3b\u4e2d\u3002\u7528\u673a\u68b0\u53ef\u89e3\u91ca\u6027\u7684\u672f\u8bed\u6765\u8bf4\uff0c\u8fd9\u53ef\u4ee5\u7b80\u6d01\u5730\u63cf\u8ff0\u4e3a\u300c\u89e3\u7801\u5668\u5411\u91cf\u5bf9\u5e94\u4e8e\u6b8b\u5dee\u6d41\u7a7a\u95f4\u4e2d\u7279\u5f81\u7684\u7ebf\u6027\u8868\u5f81\u300d\u3002<\/p>\n<p>\u4e5f\u53ef\u4ee5\u8bf4\u6709 49,512 \u7ef4\u7684\u5df2\u7f16\u7801\u8868\u5f81\u7684 SAE \u6709 49,512 \u4e2a\u7279\u5f81\u3002\u7279\u5f81\u7531\u5bf9\u5e94\u7684\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u5411\u91cf\u6784\u6210\u3002\u7f16\u7801\u5668\u5411\u91cf\u7684\u4f5c\u7528\u662f\u68c0\u6d4b\u6a21\u578b\u7684\u5185\u90e8\u6982\u5ff5\uff0c\u540c\u65f6\u6700\u5c0f\u5316\u5176\u5b83\u6982\u5ff5\u7684\u5e72\u6270\uff0c\u5c3d\u7ba1\u89e3\u7801\u5668\u5411\u91cf\u7684\u4f5c\u7528\u662f\u8868\u793a\u300c\u771f\u5b9e\u7684\u300d\u7279\u5f81\u65b9\u5411\u3002\u7814\u7a76\u8005\u7684\u5b9e\u9a8c\u53d1\u73b0\uff0c\u6bcf\u4e2a\u7279\u5f81\u7684\u7f16\u7801\u5668\u548c\u89e3\u7801\u5668\u7279\u5f81\u662f\u4e0d\u4e00\u6837\u7684\uff0c\u5e76\u4e14\u4f59\u5f26\u76f8\u4f3c\u5ea6\u7684\u4e2d\u4f4d\u6570\u4e3a 0.5\u3002\u5728\u4e0b\u56fe\u4e2d\uff0c\u4e09\u4e2a\u7ea2\u6846\u5bf9\u5e94\u4e8e\u5355\u4e2a\u7279\u5f81\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-4e34a0f1a549fab0ed3169f65988b57d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-4e34a0f1a549fab0ed3169f65988b57d.png\"><\/a><\/p>\n<p><em><sup><mark data-type=\"tech_methods\" data-id=\"df5b0394-985c-43f8-8063-9d8137501ffd\">\u7a00\u758f\u81ea\u7f16\u7801\u5668<\/mark>\u793a\u610f\u56fe\uff0c\u5176\u4e2d\u4e09\u4e2a\u7ea2\u6846\u5bf9\u5e94\u4e8e SAE \u7279\u5f81 1\uff0c\u7eff\u6846\u5bf9\u5e94\u4e8e\u7279\u5f81 4\u3002\u6bcf\u4e2a\u7279\u5f81\u90fd\u6709\u4e00\u4e2a 1&#215;4 \u7684\u7f16\u7801\u5668\u5411\u91cf\u30011&#215;1 \u7684\u7279\u5f81\u6fc0\u6d3b\u548c 1&#215;4 \u7684\u89e3\u7801\u5668\u5411\u91cf\u3002\u91cd\u5efa\u7684\u6fc0\u6d3b\u7684\u6784\u5efa\u4ec5\u4f7f\u7528\u4e86\u6765\u81ea SAE \u7279\u5f81 1 \u548c 4 \u7684\u89e3\u7801\u5668\u5411\u91cf\u3002\u5982\u679c\u7ea2\u6846\u8868\u793a\u300c\u7ea2\u989c\u8272\u300d\uff0c\u7eff\u6846\u8868\u793a\u300c\u7403\u300d\uff0c\u90a3\u4e48\u8be5\u6a21\u578b\u53ef\u80fd\u8868\u793a\u300c\u7ea2\u7403\u300d\u3002<\/sup><\/em><\/p>\n<p>\u90a3\u4e48\u6211\u4eec\u8be5\u5982\u4f55\u5f97\u77e5\u5047\u8bbe\u7684\u7279\u5f81 317 \u8868\u793a\u4ec0\u4e48\u5462\uff1f\u76ee\u524d\u800c\u8a00\uff0c\u4eba\u4eec\u7684\u5b9e\u8df5\u65b9\u6cd5\u662f\u5bfb\u627e\u80fd\u6700\u5927\u7a0b\u5ea6\u6fc0\u6d3b\u7279\u5f81\u5e76\u5bf9\u5b83\u4eec\u7684\u53ef\u89e3\u91ca\u6027\u7ed9\u51fa\u76f4\u89c9\u53cd\u5e94\u7684\u8f93\u5165\u3002\u80fd\u8ba9\u6bcf\u4e2a\u7279\u5f81\u6fc0\u6d3b\u7684\u8f93\u5165\u901a\u5e38\u662f\u53ef\u89e3\u91ca\u7684\u3002<\/p>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0cAnthropic \u5728 Claude Sonnet \u4e0a\u8bad\u7ec3\u4e86 SAE\uff0c\u7ed3\u679c\u53d1\u73b0\uff1a\u4e0e\u91d1\u95e8\u5927\u6865\u3001<mark data-type=\"concepts\" data-id=\"306453a1-dae0-4518-9204-c5d533cedbe7\">\u795e\u7ecf\u79d1\u5b66<\/mark>\u548c\u70ed\u95e8\u65c5\u6e38\u666f\u70b9\u76f8\u5173\u7684\u6587\u672c\u548c\u56fe\u50cf\u4f1a\u6fc0\u6d3b\u4e0d\u540c\u7684 SAE \u7279\u5f81\u3002\u5176\u5b83\u4e00\u4e9b\u7279\u5f81\u4f1a\u88ab\u5e76\u4e0d\u663e\u800c\u6613\u89c1\u7684\u6982\u5ff5\u6fc0\u6d3b\uff0c\u6bd4\u5982\u5728 Pythia \u4e0a\u8bad\u7ec3\u7684\u4e00\u4e2a SAE \u7684\u4e00\u4e2a\u7279\u5f81\u4f1a\u88ab\u8fd9\u6837\u7684\u6982\u5ff5\u6fc0\u6d3b\uff0c\u5373\u300c\u7528\u4e8e\u4fee\u9970\u53e5\u5b50\u4e3b\u8bed\u7684\u5173\u7cfb\u4ece\u53e5\u6216\u4ecb\u8bcd\u77ed\u8bed\u7684\u6700\u7ec8 token\u300d\u3002<\/p>\n<p>\u7531\u4e8e SAE \u89e3\u7801\u5668\u5411\u91cf\u7684\u5f62\u72b6\u4e0e LLM \u7684\u4e2d\u95f4\u6fc0\u6d3b\u4e00\u6837\uff0c\u56e0\u6b64\u53ef\u7b80\u5355\u5730\u901a\u8fc7\u5c06\u89e3\u7801\u5668\u5411\u91cf\u52a0\u5165\u5230\u6a21\u578b\u6fc0\u6d3b\u6765\u6267\u884c\u56e0\u679c\u5e72\u9884\u3002\u901a\u8fc7\u8ba9\u8be5\u89e3\u7801\u5668\u5411\u91cf\u4e58\u4ee5\u4e00\u4e2a\u6269\u5c55\u56e0\u5b50\uff0c\u53ef\u4ee5\u8c03\u6574\u8fd9\u79cd\u5e72\u9884\u7684\u5f3a\u5ea6\u3002\u5f53 Anthropic \u7814\u7a76\u8005\u5c06\u300c\u91d1\u95e8\u5927\u6865\u300dSAE \u89e3\u7801\u5668\u5411\u91cf\u6dfb\u52a0\u5230 Claude \u7684\u6fc0\u6d3b\u65f6\uff0cClaude \u4f1a\u88ab\u8feb\u5728\u6bcf\u4e2a\u54cd\u5e94\u4e2d\u90fd\u63d0\u53ca\u300c\u91d1\u95e8\u5927\u6865\u300d\u3002<\/p>\n<p>\u4e0b\u9762\u662f\u4f7f\u7528\u5047\u8bbe\u7684\u7279\u5f81 317 \u5f97\u5230\u7684\u56e0\u679c\u5e72\u9884\u7684\u53c2\u8003\u5b9e\u73b0\u3002\u7c7b\u4f3c\u4e8e\u300c\u91d1\u95e8\u5927\u6865\u300dClaude\uff0c\u8fd9\u79cd\u975e\u5e38\u7b80\u5355\u7684\u5e72\u9884\u4f1a\u8feb\u4f7f GPT-3 \u6a21\u578b\u5728\u6bcf\u4e2a\u54cd\u5e94\u4e2d\u90fd\u63d0\u53ca\u300c\u91d1\u6bdb\u72ac\u300d\u3002<\/p>\n<section>\n<pre data-lang=\"properties\"><code>def perform_intervention (model_activations_D: torch.Tensor, decoder_FD: torch.Tensor, scale: float) -&gt; torch.Tensor:<\/code>\r\n<code>    intervention_vector_D = decoder_FD [317, :]<\/code>\r\n<code>    scaled_intervention_vector_D = intervention_vector_D * scale<\/code>\r\n<code>    modified_model_activations_D = model_activations_D + scaled_intervention_vector_D<\/code>\r\n<code>\u00a0 \u00a0 return modified_model_activations_D<\/code><\/pre>\n<\/section>\n<p><strong><mark data-type=\"tech_methods\" data-id=\"df5b0394-985c-43f8-8063-9d8137501ffd\">\u7a00\u758f\u81ea\u7f16\u7801\u5668<\/mark>\u7684\u8bc4\u4f30\u96be\u9898<\/strong><\/p>\n<p>\u4f7f\u7528 SAE \u7684\u4e00\u5927\u4e3b\u8981\u96be\u9898\u662f\u8bc4\u4f30\u3002\u6211\u4eec\u53ef\u4ee5\u8bad\u7ec3<mark data-type=\"tech_methods\" data-id=\"df5b0394-985c-43f8-8063-9d8137501ffd\">\u7a00\u758f\u81ea\u7f16\u7801\u5668<\/mark>\u6765\u89e3\u91ca<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff0c\u4f46\u6211\u4eec\u6ca1\u6709\u81ea\u7136\u8bed\u8a00\u8868\u793a\u7684\u53ef\u5ea6\u91cf\u7684\u5e95\u5c42 ground truth\u3002\u76ee\u524d\u800c\u8a00\uff0c\u8bc4\u4f30\u90fd\u5f88\u4e3b\u89c2\uff0c\u57fa\u672c\u4e5f\u5c31\u662f\u300c\u6211\u4eec\u7814\u7a76\u4e00\u7cfb\u5217\u7279\u5f81\u7684\u6fc0\u6d3b\u8f93\u5165\uff0c\u7136\u540e\u51ed\u76f4\u89c9\u9610\u8ff0\u8fd9\u4e9b\u7279\u5f81\u7684\u53ef\u89e3\u91ca\u6027\u3002\u300d\u8fd9\u662f\u53ef\u89e3\u91ca\u6027\u9886\u57df\u7684\u4e3b\u8981\u9650\u5236\u3002<\/p>\n<p>\u7814\u7a76\u8005\u5df2\u7ecf\u53d1\u73b0\u4e86\u4e00\u4e9b\u4f3c\u4e4e\u4e0e\u7279\u5f81\u53ef\u89e3\u91ca\u6027\u76f8\u5bf9\u5e94\u7684\u5e38\u89c1\u4ee3\u7406\u6307\u6807\u3002\u6700\u5e38\u7528\u7684\u662f L0 \u548c Loss Recovered\u3002L0 \u662f SAE \u7684\u5df2\u7f16\u7801\u4e2d\u95f4\u8868\u5f81\u4e2d\u975e\u96f6\u5143\u7d20\u7684\u5e73\u5747\u6570\u91cf\u3002Loss Recovered \u662f\u4f7f\u7528\u91cd\u5efa\u7684\u6fc0\u6d3b\u66ff\u6362 GPT \u7684\u539f\u59cb\u6fc0\u6d3b\uff0c\u5e76\u6d4b\u91cf\u4e0d\u5b8c\u7f8e\u91cd\u5efa\u7ed3\u679c\u7684\u989d\u5916\u635f\u5931\u3002\u8fd9\u4e24\u4e2a\u6307\u6807\u901a\u5e38\u9700\u8981\u6743\u8861\u8003\u8651\uff0c\u56e0\u4e3a SAE \u53ef\u80fd\u4f1a\u4e3a\u4e86\u63d0\u5347\u7a00\u758f\u6027\u800c\u9009\u62e9\u4e00\u4e2a\u4f1a\u5bfc\u81f4\u91cd\u5efa\u51c6\u786e\u5ea6\u4e0b\u964d\u7684\u89e3\u3002<\/p>\n<p>\u5728\u6bd4\u8f83 SAE \u65f6\uff0c\u4e00\u79cd\u5e38\u7528\u65b9\u6cd5\u662f\u7ed8\u5236\u8fd9\u4e24\u4e2a\u53d8\u91cf\u7684\u56fe\u8868\uff0c\u7136\u540e\u68c0\u67e5\u5b83\u4eec\u4e4b\u95f4\u7684\u6743\u8861\u3002\u4e3a\u4e86\u5b9e\u73b0\u66f4\u597d\u7684\u6743\u8861\uff0c\u8bb8\u591a\u65b0\u7684 SAE \u65b9\u6cd5\uff08\u5982 <mark data-type=\"institutions\" data-id=\"83832d76-bfe1-42e7-ab87-9971cb43d50c\">DeepMind<\/mark> \u7684 Gated SAE \u548c OpenAI \u7684 TopK SAE\uff09\u5bf9\u7a00\u758f\u5ea6\u60e9\u7f5a\u505a\u4e86\u4fee\u6539\u3002\u4e0b\u56fe\u6765\u81ea <mark data-type=\"institutions\" data-id=\"83832d76-bfe1-42e7-ab87-9971cb43d50c\">DeepMind<\/mark> \u7684 Gated SAE \u8bba\u6587\u3002Gated SAE \u7531\u7ea2\u7ebf\u8868\u793a\uff0c\u4f4d\u4e8e\u56fe\u4e2d\u5de6\u4e0a\u65b9\uff0c\u8fd9\u8868\u660e\u5176\u5728\u8fd9\u79cd\u6743\u8861\u4e0a\u8868\u73b0\u66f4\u597d\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-9eaf8ddaced38c3e382181f78926ca3a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-9eaf8ddaced38c3e382181f78926ca3a.png\"><\/a><\/p>\n<p><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 Gated SAE L0 \u4e0e Loss Recovered<\/sup><\/em><\/p>\n<p>SAE \u7684\u5ea6\u91cf\u5b58\u5728\u591a\u4e2a\u96be\u5ea6\u5c42\u7ea7\u3002L0 \u548c Loss Recovered \u662f\u4e24\u4e2a\u4ee3\u7406\u6307\u6807\u3002\u4f46\u662f\uff0c\u5728\u8bad\u7ec3\u65f6\u6211\u4eec\u5e76\u4e0d\u4f1a\u4f7f\u7528\u5b83\u4eec\uff0c\u56e0\u4e3a L0 \u4e0d\u53ef\u5fae\u5206\uff0c\u800c\u5728 SAE \u8bad\u7ec3\u671f\u95f4\u8ba1\u7b97 Loss Recovered \u7684\u8ba1\u7b97\u6210\u672c\u975e\u5e38\u9ad8\u3002\u76f8\u53cd\uff0c\u6211\u4eec\u7684\u8bad\u7ec3\u635f\u5931\u7531\u4e00\u4e2a L1 \u60e9\u7f5a\u9879\u548c\u91cd\u5efa\u5185\u90e8\u6fc0\u6d3b\u7684\u51c6\u786e\u5ea6\u51b3\u5b9a\uff0c\u800c\u975e\u5176\u5bf9\u4e0b\u6e38\u635f\u5931\u7684\u5f71\u54cd\u3002<\/p>\n<p>\u8bad\u7ec3<mark data-type=\"concepts\" data-id=\"4c38563a-2d9b-439e-bfb4-21d209eeff3e\">\u635f\u5931\u51fd\u6570<\/mark>\u5e76\u4e0d\u4e0e\u4ee3\u7406\u6307\u6807\u76f4\u63a5\u5bf9\u5e94\uff0c\u5e76\u4e14\u4ee3\u7406\u6307\u6807\u53ea\u662f\u5bf9\u7279\u5f81\u53ef\u89e3\u91ca\u6027\u7684\u4e3b\u89c2\u8bc4\u4f30\u7684\u4ee3\u7406\u3002\u7531\u4e8e\u6211\u4eec\u7684\u771f\u6b63\u76ee\u6807\u662f\u300c\u4e86\u89e3\u6a21\u578b\u7684\u5de5\u4f5c\u65b9\u5f0f\u300d\uff0c\u4e3b\u89c2\u53ef\u89e3\u91ca\u6027\u8bc4\u4f30\u53ea\u662f\u4ee3\u7406\uff0c\u56e0\u6b64\u8fd8\u4f1a\u6709\u53e6\u4e00\u5c42\u4e0d\u5339\u914d\u3002LLM \u4e2d\u7684\u4e00\u4e9b\u91cd\u8981\u6982\u5ff5\u53ef\u80fd\u5e76\u4e0d\u5bb9\u6613\u89e3\u91ca\uff0c\u800c\u4e14\u6211\u4eec\u53ef\u80fd\u4f1a\u5728\u76f2\u76ee\u4f18\u5316\u53ef\u89e3\u91ca\u6027\u65f6\u5ffd\u89c6\u8fd9\u4e9b\u6982\u5ff5\u3002<\/p>\n<p><strong>\u603b\u7ed3<\/strong><\/p>\n<p>\u53ef\u89e3\u91ca\u6027\u9886\u57df\u8fd8\u6709\u5f88\u957f\u7684\u8def\u8981\u8d70\uff0c\u4f46 SAE \u662f\u771f\u6b63\u7684\u8fdb\u6b65\u3002SAE \u80fd\u5b9e\u73b0\u6709\u8da3\u7684\u65b0\u5e94\u7528\uff0c\u6bd4\u5982\u4e00\u79cd\u7528\u4e8e\u67e5\u627e\u300c\u91d1\u95e8\u5927\u6865\u300d\u5bfc\u5411\u5411\u91cf\uff08steering vector\uff09\u8fd9\u6837\u7684\u5bfc\u5411\u5411\u91cf\u7684\u65e0\u76d1\u7763\u65b9\u6cd5\u3002SAE \u4e5f\u80fd\u5e2e\u52a9\u6211\u4eec\u66f4\u8f7b\u677e\u5730\u67e5\u627e<mark data-type=\"tech_tasks\" data-id=\"fee178e8-ee20-42fc-8f8f-1d41c1f34e8f\">\u8bed\u8a00\u6a21\u578b<\/mark>\u4e2d\u7684\u56de\u8def\uff0c\u8fd9\u6216\u53ef\u7528\u4e8e\u79fb\u9664\u6a21\u578b\u5185\u90e8\u4e0d\u5fc5\u8981\u7684\u504f\u7f6e\u3002<\/p>\n<p>SAE \u80fd\u627e\u5230\u53ef\u89e3\u91ca\u7684\u7279\u5f81\uff08\u5373\u4fbf\u76ee\u6807\u4ec5\u4ec5\u662f\u8bc6\u522b\u6fc0\u6d3b\u4e2d\u7684\u6a21\u5f0f\uff09\uff0c\u8fd9\u4e00\u4e8b\u5b9e\u8bf4\u660e\u5b83\u4eec\u80fd\u591f\u63ed\u793a\u4e00\u4e9b\u6709\u610f\u4e49\u7684\u4e1c\u897f\u3002\u8fd8\u6709\u8bc1\u636e\u8868\u660e LLM \u786e\u5b9e\u80fd\u5b66\u4e60\u5230\u4e00\u4e9b\u6709\u610f\u4e49\u7684\u4e1c\u897f\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662f\u8bb0\u5fc6\u8868\u5c42\u7684\u7edf\u8ba1\u89c4\u5f8b\u3002<\/p>\n<p>SAE \u4e5f\u80fd\u4ee3\u8868 Anthropic \u7b49\u516c\u53f8\u66fe\u5f15\u4ee5\u4e3a\u76ee\u6807\u7684\u65e9\u671f\u91cc\u7a0b\u7891\uff0c\u5373\u300c\u7528\u4e8e<mark data-type=\"tech_methods\" data-id=\"1a0e9c5e-6502-4cd7-8683-6b5ca6c48be2\">\u673a\u5668\u5b66\u4e60<\/mark>\u6a21\u578b\u7684 MRI\uff08\u78c1\u5171\u632f\u6210\u50cf\uff09\u300d\u3002SAE \u76ee\u524d\u8fd8\u4e0d\u80fd\u63d0\u4f9b\u5b8c\u7f8e\u7684\u7406\u89e3\u80fd\u529b\uff0c\u4f46\u5374\u53ef\u7528\u4e8e\u68c0\u6d4b\u4e0d\u826f\u884c\u4e3a\u3002SAE \u548c SAE \u8bc4\u4f30\u7684\u4e3b\u8981\u6311\u6218\u5e76\u975e\u4e0d\u53ef\u514b\u670d\uff0c\u5e76\u4e14\u73b0\u5728\u5df2\u6709\u5f88\u591a\u7814\u7a76\u8005\u5728\u653b\u575a\u8fd9\u4e00\u8bfe\u9898\u3002<\/p>\n<p>\u6709\u5173<mark data-type=\"tech_methods\" data-id=\"df5b0394-985c-43f8-8063-9d8137501ffd\">\u7a00\u758f\u81ea\u7f16\u7801\u5668<\/mark>\u7684\u8fdb\u4e00\u6b65\u4ecb\u7ecd\uff0c\u53ef\u53c2\u9605 Callum McDougal \u7684 Colab \u7b14\u8bb0\u672c\uff1ahttps:\/\/www.lesswrong.com\/posts\/LnHowHgmrMbWtpkxx\/intro-to-superposition-and-sparse-autoencoders-colab<\/p>\n<p><em><sup>\u53c2\u8003\u94fe\u63a5\uff1a<\/sup><\/em><\/p>\n<p><sup><em>https:\/\/www.reddit.com\/r\/MachineLearning\/comments\/1eeihdl\/d_an_intuitive_explanation_of_sparse_autoencoders\/<\/em><\/sup><\/p>\n<p><em><sup>https:\/\/adamkarvonen.github.io\/machine_learning\/2024\/06\/11\/sae-intuitions.html<\/sup><\/em><\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-08-05-5\" target=\"_blank\">LLM\u53ef\u89e3\u91ca\u6027\u7684\u672a\u6765\u5e0c\u671b\uff1f\u7a00\u758f\u81ea\u7f16\u7801\u5668\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8fd9\u91cc\u6709\u4e00\u4efd\u76f4\u89c2\u8bf4\u660e<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:LLM\u53ef\u89e3\u91ca\u6027\u7684\u672a\u6765\u5e0c [&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":[55],"class_list":["post-34919","post","type-post","status-publish","format-standard","hentry","category-news","tag-55"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>LLM\u53ef\u89e3\u91ca\u6027\u7684\u672a\u6765\u5e0c\u671b\uff1f\u7a00\u758f\u81ea\u7f16\u7801\u5668\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8fd9\u91cc\u6709\u4e00\u4efd\u76f4\u89c2\u8bf4\u660e - \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=34919\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34919#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34919\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"LLM\u53ef\u89e3\u91ca\u6027\u7684\u672a\u6765\u5e0c\u671b\uff1f\u7a00\u758f\u81ea\u7f16\u7801\u5668\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8fd9\u91cc\u6709\u4e00\u4efd\u76f4\u89c2\u8bf4\u660e\",\"datePublished\":\"2024-12-15T02:59:10+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34919\"},\"wordCount\":242,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34919#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/08\\\/frc-9b60a79aed716aaa25e35f01a3df08ad.png\",\"keywords\":[\"\u673a\u5668\u5b66\u4e60\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=34919#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34919\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=34919\",\"name\":\"LLM\u53ef\u89e3\u91ca\u6027\u7684\u672a\u6765\u5e0c\u671b\uff1f\u7a00\u758f\u81ea\u7f16\u7801\u5668\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8fd9\u91cc\u6709\u4e00\u4efd\u76f4\u89c2\u8bf4\u660e - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34919#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34919#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/08\\\/frc-9b60a79aed716aaa25e35f01a3df08ad.png\",\"datePublished\":\"2024-12-15T02:59:10+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34919#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=34919\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34919#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/08\\\/frc-9b60a79aed716aaa25e35f01a3df08ad.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/08\\\/frc-9b60a79aed716aaa25e35f01a3df08ad.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=34919#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"LLM\u53ef\u89e3\u91ca\u6027\u7684\u672a\u6765\u5e0c\u671b\uff1f\u7a00\u758f\u81ea\u7f16\u7801\u5668\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8fd9\u91cc\u6709\u4e00\u4efd\u76f4\u89c2\u8bf4\u660e\"}]},{\"@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\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"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":"LLM\u53ef\u89e3\u91ca\u6027\u7684\u672a\u6765\u5e0c\u671b\uff1f\u7a00\u758f\u81ea\u7f16\u7801\u5668\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8fd9\u91cc\u6709\u4e00\u4efd\u76f4\u89c2\u8bf4\u660e - \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=34919","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=34919#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=34919"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"LLM\u53ef\u89e3\u91ca\u6027\u7684\u672a\u6765\u5e0c\u671b\uff1f\u7a00\u758f\u81ea\u7f16\u7801\u5668\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8fd9\u91cc\u6709\u4e00\u4efd\u76f4\u89c2\u8bf4\u660e","datePublished":"2024-12-15T02:59:10+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=34919"},"wordCount":242,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=34919#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-9b60a79aed716aaa25e35f01a3df08ad.png","keywords":["\u673a\u5668\u5b66\u4e60"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=34919#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34919","url":"https:\/\/17aitech.com\/?p=34919","name":"LLM\u53ef\u89e3\u91ca\u6027\u7684\u672a\u6765\u5e0c\u671b\uff1f\u7a00\u758f\u81ea\u7f16\u7801\u5668\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8fd9\u91cc\u6709\u4e00\u4efd\u76f4\u89c2\u8bf4\u660e - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34919#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34919#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-9b60a79aed716aaa25e35f01a3df08ad.png","datePublished":"2024-12-15T02:59:10+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34919#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34919"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34919#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-9b60a79aed716aaa25e35f01a3df08ad.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-9b60a79aed716aaa25e35f01a3df08ad.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34919#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"LLM\u53ef\u89e3\u91ca\u6027\u7684\u672a\u6765\u5e0c\u671b\uff1f\u7a00\u758f\u81ea\u7f16\u7801\u5668\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8fd9\u91cc\u6709\u4e00\u4efd\u76f4\u89c2\u8bf4\u660e"}]},{"@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":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","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\/34919","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=34919"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34919\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34919"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34919"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34919"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}