{"id":4214,"date":"2024-06-28T22:35:10","date_gmt":"2024-06-28T14:35:10","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=4214"},"modified":"2024-06-28T22:35:34","modified_gmt":"2024-06-28T14:35:34","slug":"%e7%94%a8-gpt-4-%e6%89%be%e5%87%ba-gpt-4-%e7%9a%84%e9%94%99%e8%af%af","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=4214","title":{"rendered":"\u7528 GPT-4 \u627e\u51fa GPT-4 \u7684\u9519\u8bef"},"content":{"rendered":"\n

Finding GPT-4\u2019s mistakes with GPT-4<\/strong><\/p>\n\n\n\n

CriticGPT \u662f\u4e00\u4e2a\u57fa\u4e8e GPT-4 \u7684\u6a21\u578b\uff0c\u5b83\u64b0\u5199\u4e86\u5bf9 ChatGPT \u54cd\u5e94\u7684\u8bc4\u8bba\uff0c\u4ee5\u5e2e\u52a9\u4eba\u7c7b\u8bad\u7ec3\u5e08\u5728 RLHF \u671f\u95f4\u53d1\u73b0\u9519\u8bef<\/p>\n\n\n\n

\u9605\u8bfb\u8bba\u6587<\/a><\/p>\n\n\n\n

\u6211\u4eec\u8bad\u7ec3\u4e86\u4e00\u4e2a\u57fa\u4e8e GPT-4 \u7684\u6a21\u578b\uff0c\u79f0\u4e3a CriticGPT\uff0c\u7528\u4e8e\u6355\u83b7 ChatGPT \u4ee3\u7801\u8f93\u51fa\u4e2d\u7684\u9519\u8bef\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u5f53\u4eba\u4eec\u4ece CriticGPT \u83b7\u5f97\u5e2e\u52a9\u6765\u5ba1\u67e5 ChatGPT \u4ee3\u7801\u65f6\uff0c\u4ed6\u4eec\u5728 60% \u7684\u65f6\u95f4\u91cc\u6bd4\u6ca1\u6709\u5e2e\u52a9\u7684\u4eba\u8868\u73b0\u5f97\u66f4\u597d\u3002\u6211\u4eec\u6b63\u5728\u5f00\u59cb\u5c06\u7c7b\u4f3c CriticGPT \u7684\u6a21\u578b\u96c6\u6210\u5230\u6211\u4eec\u7684 RLHF \u6807\u7b7e\u7ba1\u9053\u4e2d\uff0c\u4e3a\u6211\u4eec\u7684\u57f9\u8bad\u5e08\u63d0\u4f9b\u660e\u786e\u7684 AI \u5e2e\u52a9\u3002\u8fd9\u662f\u671d\u7740\u80fd\u591f\u8bc4\u4f30\u5148\u8fdb\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u7684\u8f93\u51fa\u8fc8\u51fa\u7684\u4e00\u6b65\uff0c\u5982\u679c\u6ca1\u6709\u66f4\u597d\u7684\u5de5\u5177\uff0c\u4eba\u4eec\u53ef\u80fd\u5f88\u96be\u5bf9\u8fd9\u4e9b\u7ed3\u679c\u8fdb\u884c\u8bc4\u5206\u3002<\/p>\n\n\n\n

\u4e3a ChatGPT \u63d0\u4f9b\u652f\u6301\u7684 GPT-4 \u7cfb\u5217\u6a21\u578b\u901a\u8fc7\u201c\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60\u201d\uff08RLHF<\/u><\/a>\uff09 \u8fdb\u884c\u8c03\u6574\uff0c\u4ee5\u63d0\u4f9b\u6709\u7528\u548c\u4e92\u52a8\u3002RLHF \u7684\u4e00\u4e2a\u5173\u952e\u90e8\u5206\u662f\u6536\u96c6\u6bd4\u8f83\uff0c\u5728\u8fd9\u4e9b\u6bd4\u8f83\u4e2d\uff0c\u88ab\u79f0\u4e3a AI \u8bad\u7ec3\u5e08\u7684\u4eba\u5bf9\u4e0d\u540c\u7684 ChatGPT \u54cd\u5e94\u8fdb\u884c\u8bc4\u5206\u3002<\/p>\n\n\n\n

\u968f\u7740\u6211\u4eec\u5728\u63a8\u7406\u548c\u6a21\u578b\u884c\u4e3a\u65b9\u9762\u7684\u8fdb\u6b65\uff0cChatGPT \u53d8\u5f97\u66f4\u52a0\u51c6\u786e\uff0c\u5b83\u7684\u9519\u8bef\u53d8\u5f97\u66f4\u52a0\u5fae\u5999\u3002\u8fd9\u53ef\u80fd\u4f1a\u4f7f AI \u8bad\u7ec3\u5e08\u5728\u51fa\u73b0\u4e0d\u51c6\u786e\u65f6\u96be\u4ee5\u53d1\u73b0\u4e0d\u51c6\u786e\u4e4b\u5904\uff0c\u4ece\u800c\u4f7f\u4e3a RLHF \u63d0\u4f9b\u652f\u6301\u7684\u6bd4\u8f83\u4efb\u52a1\u53d8\u5f97\u66f4\u52a0\u56f0\u96be\u3002\u8fd9\u662f RLHF \u7684\u4e00\u4e2a\u57fa\u672c\u5c40\u9650\u6027\uff0c\u5e76\u4e14\u968f\u7740\u6a21\u578b\u9010\u6e10\u53d8\u5f97\u6bd4\u4efb\u4f55\u53ef\u4ee5\u63d0\u4f9b\u53cd\u9988\u7684\u4eba\u66f4\u4e86\u89e3\u6a21\u578b\uff0c\u5b83\u53ef\u80fd\u4f1a\u4f7f\u8c03\u6574\u6a21\u578b\u53d8\u5f97\u8d8a\u6765\u8d8a\u56f0\u96be\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u5e2e\u52a9\u5e94\u5bf9\u8fd9\u4e00\u6311\u6218\uff0c\u6211\u4eec\u8bad\u7ec3\u00a0CriticGPT<\/strong>\u00a0\u64b0\u5199\u8bc4\u8bba\uff0c\u5f3a\u8c03 ChatGPT \u7b54\u6848\u4e2d\u7684\u4e0d\u51c6\u786e\u4e4b\u5904\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

CriticGPT \u7684\u5efa\u8bae\u5e76\u4e0d\u603b\u662f\u6b63\u786e\u7684\uff0c\u4f46\u6211\u4eec\u53d1\u73b0\uff0c\u4e0e\u6ca1\u6709 AI \u5e2e\u52a9\u76f8\u6bd4\uff0c\u5b83\u4eec\u53ef\u4ee5\u5e2e\u52a9\u57f9\u8bad\u5e08\u901a\u8fc7\u6a21\u578b\u7f16\u5199\u7684\u7b54\u6848\u53d1\u73b0\u66f4\u591a\u7684\u95ee\u9898\u3002\u6b64\u5916\uff0c\u5f53\u4eba\u4eec\u4f7f\u7528 CriticGPT \u65f6\uff0c\u4eba\u5de5\u667a\u80fd\u589e\u5f3a\u4e86\u4ed6\u4eec\u7684\u6280\u80fd\uff0c\u4ece\u800c\u6bd4\u4eba\u4eec\u5355\u72ec\u5de5\u4f5c\u65f6\u66f4\u5168\u9762\u7684\u6279\u8bc4\uff0c\u5e76\u4e14\u6bd4\u6a21\u578b\u5355\u72ec\u5de5\u4f5c\u65f6\u66f4\u5c11\u7684\u5e7b\u89c9\u9519\u8bef\u3002\u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\uff0c\u7b2c\u4e8c\u4e2a\u968f\u673a\u8bad\u7ec3\u5e08\u5728\u8d85\u8fc7 60% \u7684\u65f6\u95f4\u91cc\u66f4\u559c\u6b22\u6765\u81ea Human+CriticGPT \u56e2\u961f\u7684\u6279\u8bc4\uff0c\u800c\u4e0d\u662f\u6765\u81ea\u65e0\u52a9\u8005\u7684\u6279\u8bc4\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

CriticGPT \u5e2e\u52a9\u57f9\u8bad\u5e08\u5199\u51fa\u6bd4\u6ca1\u6709\u5e2e\u52a9\u66f4\u5168\u9762\u7684\u8bc4\u8bba\uff0c\u540c\u65f6\u4ea7\u751f\u7684\u5e7b\u89c9\u6bd4\u4ec5\u6765\u81ea\u6a21\u578b\u7684\u6279\u8bc4\u66f4\u5c11\u3002<\/p>\n\n\n\n

\u65b9\u6cd5<\/h2>

CriticGPT \u4e5f\u63a5\u53d7\u4e86 RLHF \u7684\u8bad\u7ec3\uff0c\u7c7b\u4f3c\u4e8e ChatGPT\u3002\u4f46\u4e0e ChatGPT \u4e0d\u540c\u7684\u662f\uff0c\u5b83\u770b\u5230\u4e86\u5927\u91cf\u5305\u542b\u9519\u8bef\u7684\u8f93\u5165\uff0c\u7136\u540e\u5b83\u4e0d\u5f97\u4e0d\u5bf9\u5176\u8fdb\u884c\u6279\u8bc4\u3002\u6211\u4eec\u8981\u6c42 AI \u8bad\u7ec3\u5e08\u624b\u52a8\u5c06\u8fd9\u4e9b\u9519\u8bef\u63d2\u5165\u5230 ChatGPT \u7f16\u5199\u7684\u4ee3\u7801\u4e2d\uff0c\u7136\u540e\u7f16\u5199\u793a\u4f8b\u53cd\u9988<\/span>\uff0c\u5c31\u597d\u50cf<\/span><\/i>\u4ed6\u4eec\u6293\u4f4f\u4e86\u521a\u521a\u63d2\u5165\u7684\u9519\u8bef\u4e00\u6837\u3002\u7136\u540e\uff0c\u540c\u4e00\u4e2a\u4eba\u6bd4\u8f83\u4e86\u5bf9\u4fee\u6539\u540e\u7684\u4ee3\u7801\u7684\u591a\u4e2a\u6279\u8bc4\uff0c\u8fd9\u6837\u4ed6\u4eec\u5c31\u53ef\u4ee5\u5f88\u5bb9\u6613\u5730\u5224\u65ad\u8bc4\u8bba\u4f55\u65f6\u6293\u4f4f\u4e86\u4ed6\u4eec\u63d2\u5165\u7684\u9519\u8bef\u3002\u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u7814\u7a76\u4e86 CriticGPT \u662f\u5426\u53ef\u4ee5\u6355\u83b7\u63d2\u5165\u7684\u9519\u8bef\u548c\u4e4b\u524d\u57f9\u8bad\u5e08\u6355\u83b7\u7684\u201c\u81ea\u7136\u53d1\u751f\u7684\u201dChatGPT \u9519\u8bef\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u5728 63% \u7684\u5173\u4e8e\u81ea\u7136\u53d1\u751f\u7684\u9519\u8bef\u7684\u6848\u4f8b\u4e2d\uff0c\u57f9\u8bad\u5e08\u66f4\u559c\u6b22 CriticGPT \u6279\u8bc4\u800c\u4e0d\u662f ChatGPT \u6279\u8bc4\uff0c\u90e8\u5206\u539f\u56e0\u662f\u65b0\u6279\u8bc4\u5bb6\u4ea7\u751f\u7684\u201c\u5439\u6bdb\u6c42\u75b5\u201d\uff08\u65e0\u76ca\u7684\u5c0f\u62b1\u6028\uff09\u66f4\u5c11\uff0c\u4ea7\u751f\u5e7b\u89c9\u95ee\u9898\u7684\u9891\u7387\u4e5f\u66f4\u4f4e\u3002\u6211\u4eec\u8fd8\u53d1\u73b0\uff0c\u901a\u8fc7\u5bf9\u6279\u8bc4\u5956\u52b1\u6a21\u578b\u4f7f\u7528\u989d\u5916\u7684\u6d4b\u8bd5\u65f6\u641c\u7d22\uff0c\u6211\u4eec\u53ef\u4ee5\u751f\u6210\u66f4\u957f\u3001\u66f4\u5168\u9762\u7684\u6279\u8bc4\u3002\u8fd9\u79cd\u641c\u7d22\u8fc7\u7a0b\u4f7f\u6211\u4eec\u80fd\u591f\u5e73\u8861\u5728\u4ee3\u7801\u4e2d\u5bfb\u627e\u95ee\u9898\u7684\u79ef\u6781\u7a0b\u5ea6\uff0c\u5e76\u5728\u5e7b\u89c9\u548c\u68c0\u6d4b\u5230\u7684\u9519\u8bef\u6570\u91cf\u4e4b\u95f4\u914d\u7f6e\u7cbe\u786e\u53ec\u56de\u6743\u8861\u3002\u8fd9\u610f\u5473\u7740\u6211\u4eec\u53ef\u4ee5\u751f\u6210\u5bf9 RLHF \u5c3d\u53ef\u80fd\u6709\u5e2e\u52a9\u7684\u8bc4\u8bba\u3002\u6709\u5173\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u9605\u6211\u4eec\u7684\u7814\u7a76\u8bba\u6587\u3002<\/span><\/p>

\u5c40\u9650\u6027

  • \u6211\u4eec\u5728 ChatGPT \u7684\u7b54\u6848\u4e0a\u8bad\u7ec3\u4e86 CriticGPT\uff0c\u8fd9\u4e9b\u7b54\u6848\u5f88\u77ed\u3002\u4e3a\u4e86\u76d1\u7763\u672a\u6765\u7684\u4ee3\u7406\uff0c\u6211\u4eec\u9700\u8981\u5f00\u53d1\u53ef\u4ee5\u5e2e\u52a9\u57f9\u8bad\u5e08\u7406\u89e3\u957f\u671f\u590d\u6742\u4efb\u52a1\u7684\u65b9\u6cd5\u3002\u6a21\u7279\u4ecd\u7136\u4f1a\u4ea7\u751f\u5e7b\u89c9\uff0c\u6709\u65f6\u6559\u7ec3\u5728\u770b\u5230\u8fd9\u4e9b\u5e7b\u89c9\u540e\u4f1a\u72af\u6807\u7b7e\u9519\u8bef\u3002\u6709\u65f6\uff0c\u73b0\u5b9e\u4e16\u754c\u7684\u9519\u8bef\u53ef\u80fd\u4f1a\u5206\u6563\u5728\u7b54\u6848\u7684\u8bb8\u591a\u90e8\u5206\u3002\u6211\u4eec\u7684\u5de5\u4f5c\u91cd\u70b9\u662f\u53ef\u4ee5\u5728\u4e00\u4e2a\u5730\u65b9\u6307\u51fa\u7684\u9519\u8bef\uff0c\u4f46\u5c06\u6765\u6211\u4eec\u4e5f\u9700\u8981\u89e3\u51b3\u5206\u6563\u7684\u9519\u8bef\u3002CriticGPT \u53ea\u80fd\u63d0\u4f9b\u8fd9\u4e48\u591a\u5e2e\u52a9\uff1a\u5982\u679c\u4e00\u9879\u4efb\u52a1\u6216\u54cd\u5e94\u975e\u5e38\u590d\u6742\uff0c\u5373\u4f7f\u662f\u6709\u6a21\u578b\u5e2e\u52a9\u7684\u4e13\u5bb6\u4e5f\u53ef\u80fd\u65e0\u6cd5\u6b63\u786e\u8bc4\u4f30\u5b83\u3002<\/span><\/p><\/li><\/h2>

    <\/p>

    \u540e\u7eed\u6b65\u9aa4<\/h2>

    \u4e3a\u4e86\u8c03\u6574\u65e5\u76ca\u590d\u6742\u7684\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\uff0c\u6211\u4eec\u9700\u8981\u66f4\u597d\u7684\u5de5\u5177\u3002\u5728\u6211\u4eec\u5bf9 CriticGPT \u7684\u7814\u7a76\u4e2d\uff0c\u6211\u4eec\u53d1\u73b0\u5c06 RLHF \u5e94\u7528\u4e8e GPT-4 \u6709\u671b\u5e2e\u52a9\u4eba\u7c7b\u4e3a GPT-4 \u751f\u6210\u66f4\u597d\u7684 RLHF \u6570\u636e\u3002\u6211\u4eec\u6b63\u8ba1\u5212\u8fdb\u4e00\u6b65\u6269\u5927\u8fd9\u9879\u5de5\u4f5c\u7684\u89c4\u6a21\u5e76\u5c06\u5176\u4ed8\u8bf8\u5b9e\u8df5\u3002<\/span><\/p><\/div><\/div><\/p>\n\n\n\n

    \u539f\u6587\u8fde\u63a5\uff1a\u4f7f\u7528 GPT-4 \u67e5\u627e GPT-4 \u7684\u9519\u8bef |OpenAI<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

    Finding GPT-4\u2019s mistakes with GPT-4 CriticGPT \u662f\u4e00\u4e2a\u57fa\u4e8e GPT […]<\/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":"","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":"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":""}},"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[444,445,313,289,312],"tags":[242,502,314],"class_list":["post-4214","post","type-post","status-publish","format-standard","hentry","category-ai","category-ainews","category-chatgpt","category-gpt","category-openai","tag-chatgpt","tag-criticgpt","tag-openai-api"],"views":2440,"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/4214","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4214"}],"version-history":[{"count":2,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/4214\/revisions"}],"predecessor-version":[{"id":4218,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/4214\/revisions\/4218"}],"wp:attachment":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4214"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4214"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4214"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}