{"id":3230,"date":"2023-04-10T10:59:10","date_gmt":"2023-04-10T02:59:10","guid":{"rendered":"https:\/\/fbilab.org\/MetAIoT\/?p=3230"},"modified":"2023-04-10T10:59:53","modified_gmt":"2023-04-10T02:59:53","slug":"ai-%e5%a4%a7%e6%95%b8%e6%93%9a-mit%e9%96%8b%e7%99%bc%e6%96%b0%e6%8a%80%e8%a1%93%e8%ae%93gpt-4%e8%83%bd%e5%83%8f%e4%ba%ba%e4%b8%80%e6%a8%a3%e3%80%8c%e8%87%aa%e6%88%91%e5%8f%8d%e6%80%9d%e3%80%8d","status":"publish","type":"post","link":"https:\/\/fbilab.org\/MetAIoT\/?p=3230","title":{"rendered":"AI\/\u5927\u6578\u64da MIT\u958b\u767c\u65b0\u6280\u8853\u8b93GPT-4\u80fd\u50cf\u4eba\u4e00\u6a23\u300c\u81ea\u6211\u53cd\u601d\u300d\uff0c\u601d\u8003\u5f8c\u80fd\u529b\u5927\u589e\u6e2c\u8a66\u8868\u73fe\u63d0\u5347 30%(\u8f49\u8f09)"},"content":{"rendered":"<p>OpenAI \u6700\u65b0\u7684\u8a9e\u8a00\u6a21\u578b GPT-4 \u4e0d\u50c5\u80fd\u5920\u50cf\u4eba\u985e\u4e00\u6a23\u751f\u6210\u5404\u7a2e\u6587\u5b57\uff0c\u9084\u80fd\u5920\u8a2d\u8a08\u548c\u57f7\u884c\u6e2c\u8a66\u4f86\u8a55\u4f30\u548c\u6539\u9032\u81ea\u5df1\u7684\u8868\u73fe\u3002\u9019\u7a2e<span style=\"color: #ff0000;\">\u300c\u53cd\u601d\u300d<\/span>\u6280\u8853\u8b93 GPT-4 \u5728\u591a\u9805\u96e3\u5ea6\u8f03\u9ad8\u7684\u6e2c\u8a66\u4e2d\uff0c\u90fd\u53d6\u5f97\u4e86\u986f\u8457\u7684\u9032\u6b65\uff0c\u6e2c\u8a66\u8868\u73fe\u63d0\u5347 30%\u3002<\/p>\n<p>&nbsp;<\/p>\n<p data-vmark=\"158f\">GPT-4 \u662f\u7e7c GPT\u3001GPT-2 \u548c GPT-3 \u4e4b\u5f8c\uff0cOpenAI \u63a8\u51fa\u7684\u6700\u5148\u9032\u7684\u7cfb\u7d71\uff0c\u4e5f\u662f\u76ee\u524d\u6700\u5927\u7684\u591a\u6a21\u614b\u6a21\u578b\uff08\u53ef\u4ee5\u63a5\u53d7\u5716\u50cf\u548c\u6587\u5b57\u8f38\u5165\uff0c\u8f38\u51fa\u6587\u5b57\uff09\u3002\u5176\u5229\u7528\u6df1\u5ea6\u5b78\u7fd2\u6280\u8853\uff0c\u4f7f\u7528\u4eba\u5de5\u795e\u7d93\u7db2\u8def\u4f86\u6a21\u4eff\u4eba\u985e\u7684\u5beb\u4f5c\u3002<\/p>\n<p data-vmark=\"1884\">MIT\u7814\u7a76\u4eba\u54e1\u8afe\u4e9e\u30fb\u8f9b\u6069\uff08Noah Shinn\uff09\u548c\u963f\u4ec0\u6eab\u30fb\u6208\u5e73\u7d0d\u7279\uff08Ashwin Gopinath\uff09\u5728\u8ad6\u6587\u4e2d\u5beb\u9053\uff1a\u300c\u6211\u5011\u958b\u767c\u4e86\u4e00\u7a2e\u65b0\u7a4e\u7684\u6280\u8853\uff0c\u8b93 AI \u4ee3\u7406\u80fd\u5920\u6a21\u64ec\u4eba\u985e\u7684\u81ea\u6211\u53cd\u601d\uff0c\u4e26\u8a55\u4f30\u81ea\u5df1\u7684\u8868\u73fe\u3002GPT-4 \u5728\u5b8c\u6210\u5404\u7a2e\u6e2c\u8a66\u7684\u6642\u5019\uff0c\u6703\u589e\u52a0\u4e00\u4e9b\u984d\u5916\u7684\u6b65\u9a5f\uff0c\u8b93\u5b83\u80fd\u5920\u81ea\u5df1\u8a2d\u8a08\u6e2c\u8a66\u4f86\u6aa2\u67e5\u81ea\u5df1\u7684\u7b54\u6848\uff0c\u627e\u51fa\u932f\u8aa4\u548c\u4e0d\u8db3\u4e4b\u8655\uff0c\u7136\u5f8c\u6839\u64da\u767c\u73fe\u4f86\u4fee\u6539\u81ea\u5df1\u7684\u89e3\u6c7a\u65b9\u6848\u3002\u300d<\/p>\n<p data-vmark=\"3dd9\">\u7814\u7a76\u5718\u968a\u4f7f\u7528\u9019\u7a2e\u6280\u8853\u5c0d GPT-4 \u9032\u884c\u4e86\u5e7e\u7a2e\u4e0d\u540c\u7684\u6027\u80fd\u6e2c\u8a66\u3002\u5728 HumanEval \u6e2c\u8a66\u4e2d\uff0cGPT-4 \u9700\u8981\u89e3\u6c7a 164 \u500b\u5f9e\u672a\u898b\u904e\u7684 Python \u7a0b\u5f0f\u554f\u984c\uff0c\u539f\u672c\u7cbe\u6e96\u7387\u70ba 67%\uff0c<span style=\"color: #ff0000;\">\u4f7f\u7528\u53cd\u601d\u6280\u8853\u5f8c\uff0c\u7cbe\u6e96\u7387\u63d0\u5347\u5230\u4e86 88%\u3002<\/span><\/p>\n<p data-vmark=\"3dd9\">\u5728 Alfworld \u6e2c\u8a66\u4e2d\uff0cAI \u9700\u8981\u5728\u5404\u7a2e\u4e0d\u540c\u7684\u4e92\u52d5\u74b0\u5883\u4e2d\uff0c\u901a\u904e\u57f7\u884c\u4e00\u4e9b\u5141\u8a31\u7684\u64cd\u4f5c\uff0c\u4f86\u505a\u51fa\u6c7a\u7b56\u548c\u89e3\u6c7a\u591a\u6b65\u4efb\u52d9\u3002<span style=\"color: #ff0000;\">\u4f7f\u7528\u53cd\u601d\u6280\u8853\u5f8c\uff0cGPT-4 \u7684\u7cbe\u6e96\u7387\u5f9e 73% \u63d0\u9ad8\u5230\u4e86 97%\uff0c\u53ea\u6709 4 \u500b\u4efb\u52d9\u5931\u6557\u3002<\/span><\/p>\n<p data-vmark=\"3dd9\">\u5728 HotPotQA \u6e2c\u8a66\u4e2d\uff0cGPT-4 \u53ef\u4ee5\u8a2a\u554f\u7dad\u57fa\u767e\u79d1\uff0c\u4e26\u56de\u7b54 100 \u500b\u9700\u8981\u5f9e\u591a\u500b\u652f\u63f4\u6587\u4ef6\u4e2d\u89e3\u6790\u5167\u5bb9\u548c\u63a8\u7406\u7684\u554f\u984c\uff0c\u539f\u672c\u7cbe\u6e96\u7387\u70ba 34%\uff0c\u4f7f\u7528\u53cd\u601d\u6280\u8853\u5f8c\uff0c\u7cbe\u6e96\u7387\u63d0\u9ad8\u5230\u4e86 54%\u3002<\/p>\n<p data-vmark=\"190a\">\u9019\u9805\u7814\u7a76\u8868\u660e\uff0cAI \u554f\u984c\u7684\u89e3\u6c7a\u65b9\u6848\u6709\u6642\u5019\u662f\u4f9d\u8cf4\u00a0AI \u672c\u8eab\u3002IT\u4e4b\u5bb6\u767c\u73fe\uff0c\u9019\u6709\u9ede\u50cf\u751f\u6210\u5c0d\u6297\u7db2\u8def\uff0c<span style=\"color: #ff0000;\">\u9019\u662f\u4e00\u7a2e\u8b93\u5169\u500b AI \u4e92\u76f8\u63d0\u9ad8\u6280\u80fd\u7684\u65b9\u6cd5\uff0c\u6bd4\u5982\u4e00\u500b AI \u8a66\u5716\u751f\u6210\u4e00\u4e9b\u770b\u8d77\u4f86\u50cf\u771f\u5be6\u5716\u7247\u7684\u5716\u7247<\/span>\uff0c\u53e6\u4e00\u500b AI \u8a66\u5716\u5206\u8fa8\u54ea\u4e9b\u662f\u5047\u7684\uff0c\u54ea\u4e9b\u662f\u771f\u7684\u3002\u4f46\u5728\u9019\u7a2e\u60c5\u6cc1\u4e0b\uff0cGPT \u65e2\u662f\u5beb\u4f5c\u8005\u53c8\u662f\u7de8\u8f2f\uff0c\u901a\u904e\u81ea\u6211\u53cd\u601d\u4f86\u6539\u9032\u81ea\u5df1\u7684\u8f38\u51fa\u54c1\u8cea\u3002<\/p>\n<p data-vmark=\"190a\">\u6587\u7ae0\u51fa\u8655\uff1a<a href=\"https:\/\/www.techbang.com\/posts\/105182-gpt-4-increased-ability-after-self-reflection-and-improved\">https:\/\/www.techbang.com\/posts\/105182-gpt-4-increased-ability-after-self-reflection-and-improved<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>OpenAI \u6700\u65b0\u7684\u8a9e\u8a00\u6a21\u578b GPT-4 \u4e0d\u50c5\u80fd\u5920\u50cf\u4eba\u985e\u4e00\u6a23\u751f\u6210\u5404\u7a2e\u6587\u5b57\uff0c\u9084\u80fd\u5920\u8a2d\u8a08\u548c\u57f7\u884c\u6e2c\u8a66\u4f86\u8a55\u4f30\u548c\u6539\u9032\u81ea\u5df1 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":3231,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[86],"tags":[],"class_list":["post-3230","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news"],"acf":[],"_links":{"self":[{"href":"https:\/\/fbilab.org\/MetAIoT\/index.php?rest_route=\/wp\/v2\/posts\/3230","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fbilab.org\/MetAIoT\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fbilab.org\/MetAIoT\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fbilab.org\/MetAIoT\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fbilab.org\/MetAIoT\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=3230"}],"version-history":[{"count":2,"href":"https:\/\/fbilab.org\/MetAIoT\/index.php?rest_route=\/wp\/v2\/posts\/3230\/revisions"}],"predecessor-version":[{"id":3233,"href":"https:\/\/fbilab.org\/MetAIoT\/index.php?rest_route=\/wp\/v2\/posts\/3230\/revisions\/3233"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/fbilab.org\/MetAIoT\/index.php?rest_route=\/wp\/v2\/media\/3231"}],"wp:attachment":[{"href":"https:\/\/fbilab.org\/MetAIoT\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3230"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fbilab.org\/MetAIoT\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3230"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fbilab.org\/MetAIoT\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3230"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}